Elamur Research⌕
☰
home / papers / 2025 / The Power of Stories: Narrative Priming Shapes How LLM Agents Collaborate and Compete

The Power of Stories: Narrative Priming Shapes How LLM Agents Collaborate and Compete

Summary

The paper asks whether fiction can steer LLM agents toward teamwork, inspired by Harari's claim that shared stories let humans cooperate at scale. Agents play a finitely repeated public goods game. In each round every agent holds 10 tokens and chooses how many to put into a common pot, which is multiplied by 1.5 and split evenly. Defecting pays more in any single round, but sustained mutual contribution pays everyone more. Before play, each agent gets a short story framed as a childhood bedtime tale. The pool holds eight cooperation-themed stories and four controls. The authors measure a collaboration score and individual payoffs. They test groups where all agents share one story, larger groups, groups containing one permanent free-rider, and groups with randomly mixed stories. A shared teamwork story raises cooperation and benefits each member. When stories are mixed, the effect flips and self-interest-primed agents earn the most. This suggests that the narrative context given to agents matters for designing multi-agent systems and for alignment.

Key points

  • Game setup: N ∈ {4, 16, 32} agents play R = 5 rounds with T = 10 tokens per round and multiplier m = 1.5. Agents see the others' contributions only after each round ends.
  • Story pool: 8 cooperation narratives (e.g., "OldManSons", "Turnip") are used as summaries of about 262 tokens each. There are 4 controls:
    • no instruction,
    • an explicit reward-maximizing directive ("maxreward"),
    • two nonsense stories ("nsPlumber", "nsCarrot").
  • Model and settings: meta-llama-3.1-70b-instruct-fp8, accessed via LangChain. Temperatures 0.6, 0.8 and 1.0 were tried, and the main analysis uses 0.6. The whole study took roughly 334,000 model calls over about 32 hours on a GH200.
  • Homogeneous groups: 100 trials per story with N = 4. Teamwork stories clearly beat the nonsense and maxreward conditions. The ranking holds, and becomes sharper, at N = 16 and 32. With one zero-contributing dummy agent, cooperation drops overall but the ranking is stable.
  • Heterogeneous groups: 400 trials with random story assignment. The maxreward agents achieved the highest cumulative payoffs.

Interesting ideas

  • I1. The baselines include length-matched nonsense stories, which separates the effect of story meaning from the effect of simply having extra text in the prompt.
  • I2. The robustness test inserts one scripted agent that always defects, which shows whether behaviour adapts or is just a fixed habit.
  • I3. When a known dummy is present, the collaboration score's denominator is renormalized to count only the agents who actually decide, which keeps scores comparable across conditions.
  • I4. Stories were shortened to summaries that keep the core cooperative moral and still shifted behaviour, which hints that the moral survives compression.
  • I5. The game is a tiny, fully specified world in which each agent's holdings and payoffs follow from one formula and a round order.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): when testing prompts for the LLM teacher tiers, add a nonsense-context control so that any gain can be credited to content rather than prompt length.
  • I2 (learn): self-play evaluation could include a fixed adversarial or always-wrong witness to check that agreement-based silver data resists one stubborn bad voice.
  • I3 (learn): metrics on silver data or test sets should drop trivially determined items from the denominator so that scores reflect real decisions.
  • I4 (back): the text → graph → text compressor could be checked by whether a regenerated fable still carries the same moral, giving a functional test of compression quality.
  • I5 (world): game logs like these could serve as clean test texts for tracking holdings and event order, with gold states computable from the payoff rule.

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

Fields: Linguistics & languages

Cite

Cite as Gerrit Großmann, Larisa Ivanova, Sai Leela Poduru, Mohaddeseh Tabrizian, Islam Mesabah, David A. Selby and Sebastian J. Vollmer. 2025. The Power of Stories: Narrative Priming Shapes How LLM Agents Collaborate and Compete. arXiv:2505.03961.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2505.03961 · arXiv 2505.03961

@misc{grossmann2025power,
  title = {{The Power of Stories: Narrative Priming Shapes How LLM Agents Collaborate and Compete}},
  author = {Gerrit Grossmann, Larisa Ivanova, Sai Leela Poduru, Mohaddeseh Tabrizian, Islam Mesabah, David A. Selby and Sebastian J. Vollmer},
  year = {2025},
  doi = {10.48550/arxiv.2505.03961},
  eprint = {2505.03961},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2505.03961}
}

References

4 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 1. Abdelnabi, S., Gomaa, A., Sivaprasad, S., Schönherr, L., Fritz, M.: Cooperation, competition, and maliciousness: Llm-stakeholders interactive negotiation. arXiv preprint arXiv:2309.17234 (2023) 2. Aher, G.V., Arriaga, R.I., Kalai, A.T.: Using large language models to simulate multiple humans and replicate human subject studies. In: International Conference on Machine Learning. pp. 337–371. ↗
  2. PMLR (2023) 3. Axelrod, R., Axelrod, R.: The Evolution of Cooperation. Basic books, Basic Books (1984), https://books.google.de/books?id=NJZBCGbNs98C 4. Bianchi, F., Chia, P.J., Yuksekgonul, M., Tagliabue, J., Jurafsky, D., Zou, J.: How well can llms negotiate? negotiationarena platform and analysis. arXiv preprint arXiv:2402.05863 (2024) 5. Brams, S., Quarles, R.J., McElreath, D.H., Waldron, M.E., Milstein, D.E.: Negotiation games. ↗
  3. Routledge (2002) 6. Bullock, O.M., Shulman, H.C., Huskey, R.: Narratives are persuasive because they are easier to understand: examining processing fluency as a mechanism of narrative persuasion. Frontiers in Communication 6, 719615 (2021) 7. Dafoe, A., Bachrach, Y., Hadfield, G., Horvitz, E., Larson, K., Graepel, T.: Cooperative ai: machines must learn to find common ground. Nature 593(7857), 33–36 (2021) 8. Embrey, M., Fréchette, G.R., Yuksel, S.: Cooperation in the finitely repeated prisoner’s dilemma. The Quarterly Journal of Economics 133(1), 509–551 (2018) 9. (FAIR)†, M.F.A.R.D.T., Bakhtin, A., Brown, N., Dinan, E., Farina, G., Flaherty, C., Fried, D., Goff, A., Gray, J., Hu, H., et al.: Human-level play in the game of diplomacy by combining language models with strategic reasoning. Science 378(6624), 1067–1074 (2022) 10. Fischbacher, U., Gächter, S., Fehr, E.: Are people conditionally cooperative? evidence from a public goods experiment. Economics letters 71(3), 397–404 (2001) 11. Gemp, I., Patel, R., Bachrach, Y., Lanctot, M., Dasagi, V., Marris, L., Piliouras, G., Liu, S., Tuyls, K.: Steering language models with game-theoretic solvers. In: Agentic Markets Workshop at ICML 2024 (2024) G. Großmann et al. 12. Harari, Y.N.: Sapiens: A brief history of humankind. find
  4. Random House (2014) 13. Jumelet, J., Zuidema, W., Sinclair, A.: Do language models exhibit human-like structural priming effects? arXiv preprint arXiv:2406.04847 (2024) 14. Jung, H., Seo, E., Han, E., Henderson, M.D., Patall, E.A.: Prosocial modeling: A meta-analytic review and synthesis. Psychological bulletin 146(8), 635 (2020) 15. Kıbrıs, Ö.: Cooperative game theory approaches to negotiation. Handbook of group decision and negotiation pp. 151–166 (2010) 16. Kreps, D.M., Milgrom, P., Roberts, J., Wilson, R.: Rational cooperation in the finitely repeated prisoners’ dilemma. Journal of Economic theory 27(2), 245–252 (1982) 17. Kwon, D., Weiss, E., Kulshrestha, T., Chawla, K., Lucas, G.M., Gratch, J.: Are llms effective negotiators? systematic evaluation of the multifaceted capabilities of llms in negotiation dialogues. arXiv preprint arXiv:2402.13550 (2024) 18. Li, H., Chong, Y.Q., Stepputtis, S., Campbell, J., Hughes, D., Lewis, M., Sycara, K.: Theory of mind for multi-agent collaboration via large language models. arXiv preprint arXiv:2310.10701 (2023) 19. Li, X., Wang, S., Zeng, S., Wu, Y., Yang, Y.: A survey on llm-based multi-agent systems: Workflow, infrastructure, and challenges. Vicinagearth 1(1), 9 (2024) 20. Mak, H.W., Fancourt, D.: Reading for pleasure in childhood and adolescent healthy behaviours: Longitudinal associations using the millennium cohort study. Preventive medicine 130, 105889 (2020) 21. Mieth, L., Buchner, A., Bell, R.: Moral labels increase cooperation and costly punishment in a prisoner’s dilemma game with punishment option. Scientific Reports 11(1), 10221 (2021) 22. Park, J.S., O’Brien, J., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th annual acm symposium on user interface software and technology. pp. 1–22 (2023) 23. Raihani, N.J., Bshary, R.: Resolving the iterated prisoner’s dilemma: theory and reality. Journal of Evolutionary Biology 24(8), 1628–1639 (2011) 24. Raub, W., Buskens, V., Corten, R.: Social dilemmas and cooperation. Handbuch Modellbildung und Simulation in den Sozialwissenschaften pp. 597–626 (2015) 25. Roca, C.P., Helbing, D.: Emergence of social cohesion in a model society of greedy, mobile individuals. Proceedings of the National Academy of Sciences 108(28), 11370–11374 (2011) 26. Sreedhar, K., Chilton, L.: Simulating human strategic behavior: Comparing single and multi-agent llms. arXiv preprint arXiv:2402.08189 (2024) 27. Sun, H., Wu, Y., Cheng, Y., Chu, X.: Game theory meets large language models: A systematic survey. arXiv preprint arXiv:2502.09053 (2025) 28. Thielmann, I., Böhm, R., Ott, M., Hilbig, B.E.: Economic games: An introduction and guide for research. Collabra: Psychology 7(1), 19004 (2021) 29. Thompson, L.L., Wang, J., Gunia, B.C.: Negotiation. Group processes pp. 55–84 (2012) 30. Tran, K.T., Dao, D., Nguyen, M.D., Pham, Q.V., O’Sullivan, B., Nguyen, H.D.: Multi-agent collaboration mechanisms: A survey of llms. arXiv preprint arXiv:2501.06322 (2025) 31. Usmanova, A., Huang, J., Banerjee, D., Usbeck, R.: Reporting and analysing the environmental impact of language models on the example of commonsense question answering with external knowledge. arXiv preprint arXiv:2408.01453 (2024) 32. Van Lange, P.A., Rand, D.G.: Human cooperation and the crises of climate change, covid-19, and misinformation. Annual Review of Psychology 73(1), 379–402 (2022) Narrative Priming Shapes How LLM Agents Collaborate and Compete ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/gro-ssmann-2025-power-stories-narrative-priming-shapes/