Elamur Research⌕
☰
home / papers / 2026 / Communication Enables Cooperation in LLM Agents: A Comparison with Curriculum-Based Approaches

Communication Enables Cooperation in LLM Agents: A Comparison with Curriculum-Based Approaches

Paper · 2026 · Hachem Madmoun and Salem Lahlou

Summary

The paper examines how to make LLM agents cooperate in social dilemmas. It compares two interventions: an explicit communication channel and in-context curriculum learning. Four instruction-tuned models (Mixtral-8x22B, Qwen2.5-72B, Llama-3.3-70B, DeepSeek-V3) played classic game-theory scenarios through the DeepInfra API.

In a 4-player Stag Hunt, letting agents send one word before acting raised cooperation in mixed-model groups from zero to 48.3%. The curriculum had agents work through progressively harder games before a 10-round Iterated Public Goods Game with Punishment (IPGG+P). Between stages, Claude Opus 4.1 summarized strategic lessons for the agents.

Every curriculum variant scored below a no-training control, and the full sequence scored 27.4% lower. Reasoning traces showed that agents took defection lessons from short games into settings where cooperation pays off. The authors call this "learned pessimism." Replacing the lessons with neutral text removed most of the effect.

The study suggests that cheap talk is a more dependable coordination tool than experience-based prompting. It also shows that the order of games and the content of lessons can backfire.

Key points

  • Stag Hunt results: Heterogeneous groups went from 0% to 48.3% cooperation with communication. Same-family coalitions already reached 52.2% without communication and 50.0% with it. Communication also removed payoff variance in coalitions (17.9 ± 11.8 → 30.0 ± 0.0).
  • Curriculum conditions: There were 30 trials per condition (29 for the full curriculum). The control averaged 211.7 tokens. Relative to the control, the Direct Precursor scored −6.0%, the Scrambled order −14.0%, and the Full Curriculum (2-Player IPD → N-Player IPD → 3-round IPGG → IPGG+P) −27.4%, at 153.6 tokens.
  • Communication in IPGG+P:
    • At a 1.6x multiplier, contributions rose from 48% to 71%, but payoffs fell from 184.4 to 127.5 because signalling cooperators were exploited.
    • At a 4.0x multiplier, contributions reached 100% and payoffs rose from 457.9 to 480.0.
  • Neutral Lesson Ablation: Learned pessimism appeared in 62% of Full Curriculum traces versus under 5% with neutral lessons. Generic or role-play reasoning appeared in 20% versus 83%. Neutral lessons were also linked to a 63% payoff gain.
  • Validation and limitations: The authors report that GPT-4o and o1-preview showed the same communication pattern. The authors acknowledge several limitations:
    • The curriculum placed defection-equilibrium games first.
    • Human-written lessons were not tested.
    • Only in-context learning was used, with no fine-tuning.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Mohamed bin Zayed University of Artificial IntelligenceAbu Dhabi · United Arab Emirates124 works · 5,676 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSociology and Political Science · 577 worksDecision Sciences · 209 worksManagement Science and Operations Research · 108 works

Cite

Cite as Hachem Madmoun and Salem Lahlou. 2026. Communication Enables Cooperation in LLM Agents: A Comparison with Curriculum-Based Approaches. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-short.23

@inproceedings{madmoun2026communication,
  title = {{Communication Enables Cooperation in LLM Agents: A Comparison with Curriculum-Based Approaches}},
  author = {Hachem Madmoun and Salem Lahlou},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-short.23},
  url = {https://doi.org/10.18653/v1/2026.eacl-short.23}
}

References

36 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Anjali Chaudhary. 2025. find
  2. Top LLM Trends 2025: What’s the Future of LLMs. find
  3. Anthropic. 2024. Introducing Claude 4. find
  4. Robert Axelrod and William D Hamilton. 1981. The evolution of cooperation. science, 211(4489):1390– 1396. find
  5. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48.
  6. Colin Camerer. 2003. Behavioral game theory: Experiments in strategic interaction. Princeton university press. find
  7. Anthony Costarelli, Mat Allen, Roman Hauksson, Grace Sodunke, Suhas Hariharan, Carlson Cheng, Wenjie Li, Joshua Clymer, and Arjun Yadav. 2024. Gamebench: Evaluating strategic reasoning abilities of llm agents. arXiv preprint arXiv: 2406.06613.
  8. Joseph Farrell. 1987. Cheap talk, coordination, and entry. The RAND Journal of Economics, pages 34– 39. find
  9. Ernst Fehr and Simon Gächter. 2002. Altruistic punishment in humans. Nature, 415(6868):137–140. find
  10. Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. 2023. Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems, 36:70757–70798. find
  11. Drew Fudenberg and Jean Tirole. 1991. Game theory. MIT press. find
  12. Anna Alexandra Grigoryan. 2024. How LLM Agents Play the Prisoner’s Dilemma. Medium. Accessed: 2025-08-22. find
  13. Sanford J Grossman and Joseph E Stiglitz. 1976. Information and competitive price systems. The American economic review, 66(2):246–253. find
  14. Werner Güth, Rolf Schmittberger, and Bernd Schwarze. 1982. An experimental analysis of ultimatum bargaining. Journal of economic behavior & organization, 3(4):367–388. find
  15. Salem Lahlou, Abdalgader Abubaker, and Hakim Hacid. 2025. Port: Preference optimization on reasoning traces. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 10989… find
  16. Nunzio Lorè and Babak Heydari. 2023. Strategic behavior of large language models: Game structure vs. contextual framing. arXiv preprint arXiv:2309.05898. — [arXiv:2309.05898](https://arxiv.org/abs/2309.05898) find
  17. Tambet Matiisen, Avital Oliver, Taco Cohen, and John Schulman. 2019. Teacher–student curriculum learning. IEEE transactions on neural networks and learning systems, 31(9):3732–3740. find
  18. Richard D McKelvey and Thomas R Palfrey. 1992. An experimental study of the centipede game. Econometrica: Journal of the Econometric Society, pages 803–836. find
  19. Juanjuan Meng. 2024. Ai emerges as the frontier in behavioral science. Proceedings of the National Academy of Sciences, 121(10):e2401336121. find
  20. John F. Nash, Jr. 1951. Non-cooperative games. Annals of Mathematics, pages 286–295. find
  21. Martin J Osborne. 2004. An introduction to game theory, volume 3. Oxford university press. find
  22. Kenneth Payne and Baptiste Alloui-Cros. 2025. Strategic Intelligence in Large Language Models: Evidence from evolutionary Game Theory. arXiv preprint arXiv: 2507.02618. — [arXiv:2507.02618](https://arxiv.org/abs/2507.02618) find
  23. Brian Skyrms. 2004. The stag hunt and the evolution of social structure. Cambridge University Press. find
  24. Lewin Sun and 1 others. 2025. Game Theory Meets Large Language Models: A Systematic Survey with Taxonomy and New Frontiers. arXiv preprint arXiv:2502.09053. — [arXiv:2502.09053](https://arxiv.org/abs/2502.09053) find
  25. Jen tse Huang, Eric John Li, Man Ho LAM, Tian Liang, Wenxuan Wang, Youliang Yuan, Wenxiang Jiao, Xing Wang, Zhaopeng Tu, and Michael Lyu. 2025. Competing large language models in multi-agent gaming environments. In The Thirteenth International Conference on Learning Representations. find
  26. Haochuan Wang, Xiachong Feng, Lei Li, Yu Guo, Zhanyue Qin, Dianbo Sui, and Lingpeng Kong. 2024. Tmgbench: A systematic game benchmark for evaluating strategic reasoning abilities of llms. find
  27. Jingru Jia, Zehua Yuan, Junhao Pan, Paul E. McNamara, and Deming Chen. 2025. Large language model strategic reasoning evaluation through behavioral game theory. arXiv preprint arXiv: 2502.20432. — [arXiv:2502.20432](https://arxiv.org/abs/2502.20432) find
  28. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  29. Daniel Kahneman and Amos Tversky. 1979. Prospect theory: An analysis of decision under risk. Econometrica, 47(2):263–291. find
  30. Lucas Willems, Salem Lahlou, and Yoshua Bengio. 2020. Mastering rate based curriculum learning. arXiv preprint arXiv:2008.06456. — [arXiv:2008.06456](https://arxiv.org/abs/2008.06456) find
  31. Michael Wooldridge. 2009. An introduction to multiagent systems. John wiley & sons. find
  32. Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, and 1 others. 2024. Autogen: Enabling next-gen llm applications via multi-agent conversations. In First Conference on Language Modeling. find
  33. Lin Xu, Zhiyuan Hu, Daquan Zhou, Hongyu Ren, Zhen Dong, Kurt Keutzer, See Kiong Ng, and Jiashi Feng. 2024. Magic: Investigation of large language model powered multi-agent in cognition, adaptability, rationality and collaboration. In Proceedings of the 2024 Conference on Empirical Methods in Natural… find
  34. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822. find
  35. Guangran Zeng and 1 others. 2023. ALYMPICS: LLM Agents Meet Game Theory. arXiv preprint arXiv:2311.03220.
  36. Kunlun Zhu, Hongyi Du, Zhaochen Hong, Xiaocheng Yang, Shuyi Guo, Zhe Wang, Zhenhailong Wang, Cheng Qian, Xiangru Tang, Heng Ji, and Jiaxuan You. 2025. Multiagentbench: Evaluating the collaboration and competition of llm agents. arXiv preprint arXiv: 2503.01935. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/madmoun-2026-communication-enables-cooperation-llm-agents/