{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,7]],"date-time":"2026-09-07T14:26:52Z","timestamp":1788791212739,"version":"build-2803163510"},"reference-count":33,"publisher":"Frontiers Media SA","license":[{"start":{"date-parts":[[2026,7,10]],"date-time":"2026-07-10T00:00:00Z","timestamp":1783641600000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"DOI":"10.13039\/100016172","name":"Luxembourg Institute of Science and Technology","doi-asserted-by":"publisher","award":["ADIALab-MAST"],"award-info":[{"award-number":["ADIALab-MAST"]}],"id":[{"id":"10.13039\/100016172","id-type":"DOI","asserted-by":"publisher"}]},{"DOI":"10.13039\/100016172","name":"Luxembourg Institute of Science and Technology","doi-asserted-by":"publisher","award":["LLMs4EU"],"award-info":[{"award-number":["LLMs4EU"]}],"id":[{"id":"10.13039\/100016172","id-type":"DOI","asserted-by":"publisher"}]},{"DOI":"10.13039\/501100006433","name":"Barcelona Supercomputing Center","doi-asserted-by":"publisher","award":["TIFON"],"award-info":[{"award-number":["TIFON"]}],"id":[{"id":"10.13039\/501100006433","id-type":"DOI","asserted-by":"publisher"}]}],"content-domain":{"domain":["frontiersin.org"],"crossmark-restriction":true},"short-container-title":["Front. Artif. Intell."],"abstract":"<jats:sec>\n                    <jats:title>Introduction<\/jats:title>\n                    <jats:p>Large language models (LLMs) can generate persuasive narratives that shift agent behavior in multi-agent systems, but deploying raw, unstructured text as an influence mechanism offers no formal guarantees on effectiveness, interpretability, or controllability.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Methods<\/jats:title>\n                    <jats:p>\n                      We introduce the LLM Influence Compiler (LIC), a solver\u2013critic pipeline that compiles natural-language cooperation directives into structured, parameterized influence policies defined over a five-field schema: network targeting, narrative theme, intensity, deployment timing, and compiler confidence. Each compiled policy is diffused through a network of numerical agents via an exposure model incorporating fatigue decay, susceptibility heterogeneity, and backlash. Evaluation spans nine controlled experiment blocks comprising more than 200 simulation runs across four topologies (Barab\u00e1si\u2013Albert, small-world, Erd\u0151s\u2013R\u00e9nyi, modular SBM), four network sizes (\n                      <jats:italic>n<\/jats:italic>\n                      \u2208 {80, 160, 320, 640}), five LLM backbones, and two non-stationary perturbations.\n                    <\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Results<\/jats:title>\n                    <jats:p>\n                      Compiled policies raise the mean cooperation rate to 0.826 \u00b1 0.010 (95% bootstrap CI [0.819, 0.834]), an 8.6% relative improvement over the unstructured baseline (0.760 \u00b1 0.005, Mann\u2013Whitney\n                      <jats:italic>p<\/jats:italic>\n                      <jats:sub>Bonf<\/jats:sub>\n                      = 0.040, Cohen's\n                      <jats:italic>d<\/jats:italic>\n                      = 7.90). A controlled decomposition attributes the gain to network-aware targeting (\n                      <jats:italic>ca<\/jats:italic>\n                      . +4.5 pp), dose calibration (+1.9 pp), and an intervention floor (+1.5 pp); the residual contribution of full LLM compilation over a hand-coded rule with the same parameters is statistically indistinguishable from zero under stationary conditions. Under structural non-stationarity (mid-simulation graph rewire), the LLM-mediated re-deployment architecture significantly outperforms a frozen rule (paired t-test\n                      <jats:italic>p<\/jats:italic>\n                      &amp;lt; 0.001, paired Cohen's\n                      <jats:italic>d<\/jats:italic>\n                      = 2.27, 8\/8 seeds). Adversarial stress testing confirms that the critic correctly flags 20\/20 risky policies as\n                      <jats:sc>high<\/jats:sc>\n                      risk and rejects them, while passing a moderate-baseline policy at\n                      <jats:sc>medium<\/jats:sc>\n                      risk in all 5\/5 trials.\n                    <\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Discussion<\/jats:title>\n                    <jats:p>The compiler abstraction converts an opaque generative process into a decomposable, auditable policy object whose components can be independently attributed, compared, and governed, laying the groundwork for constitutional oversight of LLM-mediated influence in artificial societies.<\/jats:p>\n                  <\/jats:sec>","DOI":"10.3389\/frai.2026.1820827","type":"journal-article","created":{"date-parts":[[2026,7,10]],"date-time":"2026-07-10T05:39:05Z","timestamp":1783661945000},"update-policy":"https:\/\/doi.org\/10.3389\/crossmark-policy","source":"Crossref","is-referenced-by-count":2,"title":["From LLM narratives to parameterized cooperation policies in multi-agent systems"],"prefix":"10.3389","volume":"9","author":[{"given":"J.","family":"de Curt\u00f2","sequence":"first","affiliation":[{"name":"Department of Computer Applications in Science & Engineering, BARCELONA Supercomputing Center","place":["Barcelona, Spain"]},{"name":"Escuela T\u00e9cnica Superior de Ingenier\u00eda (ICAI), Universidad Pontificia Comillas","place":["Madrid, Spain"]},{"name":"Estudis d\u2018Inform\u00e0tica, Multim\u00e8dia i Telecomunicaci\u00f3, Universitat Oberta de Catalunya","place":["Barcelona, Spain"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"I.","family":"de Zarz\u00e0","sequence":"additional","affiliation":[{"name":"Estudis d\u2018Inform\u00e0tica, Multim\u00e8dia i Telecomunicaci\u00f3, Universitat Oberta de Catalunya","place":["Barcelona, Spain"]},{"name":"Human Centered AI, Data & Software, LUXEMBOURG Institute of Science and Technology","place":["Esch-sur-Alzette, Luxembourg"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Jordi","family":"Cabot","sequence":"additional","affiliation":[{"name":"Human Centered AI, Data & Software, LUXEMBOURG Institute of Science and Technology","place":["Esch-sur-Alzette, Luxembourg"]},{"name":"Interdisciplinary Centre for Security, Reliability and Trust (SNT), Universit\u00e9 du Luxembourg","place":["Esch-sur-Alzette, Luxembourg"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Pietro","family":"Manzoni","sequence":"additional","affiliation":[{"name":"Departamento de Inform\u00e1tica de Sistemas y Computadores, Universitat Polit\u00e8cnica de Val\u00e8ncia","place":["Val\u00e8ncia, Spain"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Carlos T.","family":"Calafate","sequence":"additional","affiliation":[{"name":"Departamento de Inform\u00e1tica de Sistemas y Computadores, Universitat Polit\u00e8cnica de Val\u00e8ncia","place":["Val\u00e8ncia, Spain"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1965","published-online":{"date-parts":[[2026,7,10]]},"reference":[{"key":"B1","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2305.16867","article-title":"Playing repeated games with large language models","author":"Akata","year":"2023","journal-title":"arXiv"},{"key":"B2","doi-asserted-by":"publisher","first-page":"49","DOI":"10.1016\/j.ins.2022.01.072","article-title":"Biased opinion dynamics: when the devil is in the details","volume":"593","author":"Anagnostopoulos","year":"2022","journal-title":"Inf. Sci"},{"key":"B3","doi-asserted-by":"publisher","first-page":"509","DOI":"10.1126\/science.286.5439.509","article-title":"Emergence of scaling in random networks","volume":"286","author":"Barab\u00e1si","year":"1999","journal-title":"Science"},{"key":"B4","volume-title":"Psychological Reactance: A Theory of Freedom and Control","author":"Brehm","year":"2013"},{"key":"B5","doi-asserted-by":"publisher","first-page":"44330","DOI":"10.1109\/ACCESS.2025.3548451","article-title":"LLM-driven social influence for cooperative behavior in multi-agent systems","volume":"13","author":"de Curt\u00f2","year":"2025","journal-title":"IEEE Access"},{"key":"B6","doi-asserted-by":"publisher","first-page":"2722","DOI":"10.3390\/electronics12122722","article-title":"Emergent cooperation and strategy adaptation in multi-agent systems: an extended coevolutionary theory with LLMs","volume":"12","author":"de Zarz\u00e0","year":"2023","journal-title":"Electronics"},{"key":"B7","doi-asserted-by":"publisher","first-page":"041101","DOI":"10.1063\/5.0201968","article-title":"Evolution of cooperation on reinforcement-learning driven-adaptive networks","volume":"34","author":"Du","year":"2024","journal-title":"Chaos"},{"key":"B8","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2305.14325","article-title":"Improving factuality and reasoning in language models through multiagent debate","author":"Du","year":"2024","journal-title":"arXiv"},{"key":"B9","doi-asserted-by":"publisher","first-page":"325","DOI":"10.1080\/01972240490507974","article-title":"The concept of information overload: a review of literature from organization science, accounting, marketing, MIS, and related disciplines","volume":"20","author":"Eppler","year":"2004","journal-title":"Inf. Soc"},{"key":"B10","doi-asserted-by":"publisher","first-page":"17960","DOI":"10.1609\/aaai.v38i16.29751","article-title":"Can large language models serve as rational players in game theory? A systematic analysis","volume":"38","author":"Fan","year":"2024","journal-title":"Proc. AAAI Conf. Artif. Intell"},{"key":"B11","doi-asserted-by":"publisher","author":"Fontana","year":"2024","DOI":"10.48550\/arXiv.2406.13605"},{"key":"B12","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2307.14984","article-title":"S3: Social-network simulation system with large language model-empowered agents","author":"Gao","year":"2023","journal-title":"arXiv"},{"key":"B13","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2402.01680","article-title":"Large language model based multi-agents: a survey of progress and challenges","author":"Guo","year":"2024","journal-title":"arXiv"},{"key":"B14","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2310.02172","article-title":"Lyfe agents: Generative agents for low-cost real-time social interactions","author":"Kaiya","year":"2023","journal-title":"arXiv"},{"key":"B15","doi-asserted-by":"crossref","first-page":"137","DOI":"10.1145\/956750.956769","article-title":"\u201cMaximizing the spread of influence through a social network,\u201d","volume-title":"Proceedings of the ninth ACM SIGKDD international conference on knowledge discovery and data mining","author":"Kempe","year":"2003"},{"key":"B16","doi-asserted-by":"crossref","DOI":"10.4324\/9781410609816","volume-title":"Resistance and Persuasion","author":"Knowles","year":"2004"},{"key":"B17","author":"Li","year":"2024"},{"key":"B18","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2305.19118","article-title":"Encouraging divergent thinking in large language models through multi-agent debate","author":"Liang","year":"2024","journal-title":"arXiv"},{"key":"B19","doi-asserted-by":"publisher","first-page":"114799","DOI":"10.1016\/j.chaos.2024.114799","article-title":"Coevolution of cognition and cooperation in structured populations under reinforcement learning","volume":"182","author":"Mastrandrea","year":"2024","journal-title":"Chaos Solitons Fractals"},{"key":"B20","doi-asserted-by":"publisher","first-page":"280","DOI":"10.1145\/3800683","article-title":"From individual to society: a survey on social simulation driven by large language model-based agents","volume":"58","author":"Mou","year":"2026","journal-title":"ACM Comput. Surv."},{"key":"B21","doi-asserted-by":"publisher","first-page":"167","DOI":"10.1137\/S003614450342480","article-title":"The structure and function of complex networks","volume":"45","author":"Newman","year":"2003","journal-title":"SIAM Rev"},{"key":"B22","doi-asserted-by":"publisher","first-page":"1560","DOI":"10.1126\/science.1133755","article-title":"Five rules for the evolution of cooperation","volume":"314","author":"Nowak","year":"2006","journal-title":"Science"},{"key":"B23","doi-asserted-by":"publisher","first-page":"502","DOI":"10.1038\/nature04605","article-title":"A simple rule for the evolution of cooperation on graphs and social networks","volume":"441","author":"Ohtsuki","year":"2006","journal-title":"Nature"},{"key":"B24","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1145\/3586183.3606763","article-title":"\u201cGenerative agents: Interactive simulacra of human behavior,\u201d","volume-title":"Proceedings of the 36th annual ACM symposium on user interface software and technology","author":"Park","year":"2023"},{"key":"B25","doi-asserted-by":"publisher","first-page":"1","DOI":"10.1016\/j.physrep.2017.05.004","article-title":"Statistical physics of human cooperation","volume":"687","author":"Perc","year":"2017","journal-title":"Phys. Rep"},{"key":"B26","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2404.16698","article-title":"Cooperate or collapse: emergence of sustainable cooperation in a society of LLM agents","author":"Piatti","year":"2024","journal-title":"arXiv"},{"key":"B27","doi-asserted-by":"publisher","first-page":"098104","DOI":"10.1103\/PhysRevLett.95.098104","article-title":"Scale-free networks provide a unifying framework for the emergence of cooperation","volume":"95","author":"Santos","year":"2005","journal-title":"Phys. Rev. Lett"},{"key":"B28","doi-asserted-by":"publisher","first-page":"1151","DOI":"10.1080\/1369118X.2017.1305427","article-title":"Too much information? Predictors of information overload in the context of online news exposure","volume":"21","author":"Schmitt","year":"2018","journal-title":"Inf. Commun. Soc"},{"key":"B29","doi-asserted-by":"publisher","first-page":"97","DOI":"10.1016\/j.physrep.2007.04.004","article-title":"Evolutionary games on graphs","volume":"446","author":"Szab\u00f3","year":"2007","journal-title":"Phys. Rep"},{"key":"B30","doi-asserted-by":"publisher","first-page":"441","DOI":"10.1086\/518527","article-title":"Influentials, networks, and public opinion formation","volume":"34","author":"Watts","year":"2007","journal-title":"J. Consumer Res"},{"key":"B31","doi-asserted-by":"publisher","first-page":"440","DOI":"10.1038\/30918","article-title":"Collective dynamics of \u201csmall-world\u201d networks","volume":"393","author":"Watts","year":"1998","journal-title":"Nature"},{"key":"B32","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2309.07864","article-title":"The rise and potential of large language model based agents: a survey","author":"Xi","year":"2023","journal-title":"arXiv"},{"key":"B33","doi-asserted-by":"publisher","first-page":"4","DOI":"10.1109\/MCAS.2021.3118100","article-title":"Analysis, prediction, and control of epidemics: a survey from scalar to dynamic network models","volume":"21","author":"Zino","year":"2021","journal-title":"IEEE Circuits Syst. Mag"}],"container-title":["Frontiers in Artificial Intelligence"],"original-title":[],"link":[{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/frai.2026.1820827\/full","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,7,10]],"date-time":"2026-07-10T05:39:09Z","timestamp":1783661949000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/frai.2026.1820827\/full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,7,10]]},"references-count":33,"alternative-id":["10.3389\/frai.2026.1820827"],"URL":"https:\/\/doi.org\/10.3389\/frai.2026.1820827","relation":{},"ISSN":["2624-8212"],"issn-type":[{"value":"2624-8212","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,7,10]]},"article-number":"1820827"}}