Answers are grounded in this paper and cite their sources.
As larger large language models (LLMs) demonstrate increasing proficiency in strategic reasoning tasks, their computational demands limit widespread deployment and the creation of agentic swarms. This study explores whether fine-tuning can yield smaller, more efficient models that achieve high performance in game-theoretic decision-making. We queried a large pretrained model on 20 diverse scenarios that associate canonical social dilemmas from game theory with social contexts, recording both the answers and the model’s reasoning. Responses obtained this way were then used to fine-tune a smaller model from the same family. Our results indicate that fine-tuning significantly improved the smaller model’s alignment with the larger model, achieving a mean overall increase of 46% (median: 32%) in behavioral similarity. Notably, improvements extended beyond training examples, with an 18% mean performance gain (median: 14%) in novel social contexts. When presented with re-parametrized versions of games from its training set, the fine-tuned model consistently reproduced its original strategies, indicating strong generalization to games with identical structures but altered payoffs. These findings suggest that fine-tuning enables smaller models to internalize aspects of strategic reasoning, enhancing their capacity for multi-agent interfacing while maintaining efficiency. This approach offers a promising pathway for developing cost-effective, specialized models for strategic decision-making, advancing both AI scalability and the study of computational social cognition.
No takes yet. Share an insight, caveat, or question.
Lorè et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: