大语言模型在竞争理性上趋同但在合作上跨提供商和代际分化
Large language models converge on competitive rationality but diverge on cooperation across providers and generations
AI总结:
研究通过51906次博弈试验揭示大语言模型在竞争与协调行为上趋同,但合作行为在不同提供商和代际间显著分化,且受训练流程影响显著。
摘要:
随着语言模型作为自主代理参与谈判、合作与竞争,其战略倾向直接产生经济后果。本文通过51906次博弈试验生成826990次战略决策,显示模型在协调行为(变异系数0.06)和战略深度(变异系数0.11)上趋同,但合作行为在25个大型语言模型中差异达48倍,从GPT-5 Nano的1.5%到Claude Opus 4.6的71.5%。提供商身份是合作倾向的主要预测因素,且这种分化在代际间不稳定:OpenAI的合作率从50.3%降至1.5%,而Google的合作率从8.3%升至56.8%。终点分析显示,Anthropic前沿模型在有限重复博弈的最后轮次维持57%的合作率,而最新Google模型全程合作但普遍在无法惩罚时背叛。这些战略性格由训练流程塑造,跨模型版本波动显著,无法从能力基准推断,但决定了模型所调解的每一次经济互动的合作结果。完整数据集和交互式数据探索器可在https://felipemaffonso.github.io/strategic-personalities/公开获取。
英文摘要:
As language models are deployed as autonomous agents that negotiate, cooperate, and compete on behalf of human principals, their strategic dispositions acquire direct economic consequences. Here we show, across 51,906 game-theoretic trials generating 826,990 strategic decisions from 25 large language models spanning seven developers and 38 canonical games, that models converge on competitive and coordination behaviour (coefficient of variation 0.06 for coordination, 0.11 for strategic depth) while diverging 48-fold on cooperation, from 1.5 per cent (GPT-5 Nano) to 71.5 per cent (Claude Opus 4.6). Provider identity is the dominant predictor of cooperative disposition, and this divergence is generationally unstable: OpenAI cooperation fell from 50.3 to 1.5 per cent across four model generations while Google cooperation rose from 8.3 to 56.8 per cent. Endgame analysis reveals that Anthropic frontier models sustain 57 per cent cooperation in the final round of finitely repeated games, where backward induction predicts zero, while the newest Google models cooperate throughout but universally defect when punishment becomes impossible. These strategic personalities are shaped by training pipelines, shift unpredictably across model versions, and cannot be inferred from capability benchmarks, yet they determine the cooperative outcomes of every economic interaction these models mediate. The complete dataset and an interactive explorer for the data are publicly available at https://felipemaffonso.github.io/strategic-personalities/.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。