基于演化博弈的生成式人工智能幻觉应对分析

doi:10.16182/j.issn1004731x.joss.25-0996

摘要/Abstract

摘要：

生成式人工智能尤其是大语言模型的广泛应用引发了“幻觉”这一社会风险。为解决现有研究主要聚焦于技术层面的幻觉缓解机制或政策层面的监管框架设计，缺乏对“大模型-用户-监管者”三方主体在有限理性条件下策略互动演化逻辑的系统性理论阐释，将演化博弈理论引入生成式人工智能治理领域，构建了融合大模型诚实性策略、用户反馈行为、监管者干预的三方动态博弈模型，揭示了多元主体在成本-收益权衡下的策略选择动态演化路径及其稳定性条件。结果表明：系统在合理参数条件下可收敛至“大模型诚实回答-用户积极反馈-监管者积极监管”的理想均衡；用户积极反馈初始意愿通过双重信号效应同步加速大模型诚实化进程与监管响应强度；激励机制呈现非对称敏感性，用户对正向激励最为敏感，监管惩罚对模型合规形成刚性约束，协同收益则在长期发挥稳定作用。

关键词: 生成式人工智能, 大语言模型, AI幻觉, 演化博弈, 协同治理

Abstract:

The accelerated deployment of generative artificial intelligence, particularly large language models, has amplified the social risks of hallucinations, posing systemic threats to the credibility of the information ecosystem, the effectiveness of users’ cognitive decision-making, and the governance security in the public domain. Research primarily focuses on hallucination mitigation mechanisms at the technical level or the design of regulatory frameworks at the policy level, lacking a systematic theoretical analysis of the evolutionary logic of strategic interactions among the “large language models, users, and regulators” under conditions of bounded rationality. By introducing evolutionary game theory into the field of generative artificial intelligence governance, a tripartite dynamic game model integrating the honesty strategies of large language models, user feedback behaviors, and regulatory interventions was constructed. This model revealed the dynamic evolutionary paths of strategy selection and their stability conditions for multiple actors under cost-benefit trade-offs. Research shows that under reasonable parameters, the system can converge to the optimal equilibrium of “honest responding of large language models, active feedback of users, and proactive oversight of regulators”. The initial willingness of users to provide positive feedback accelerates both the honesty process of large language models and the intensity of regulatory responses simultaneously through the dual signal effect. Incentive mechanisms exhibit asymmetric sensitivity: Users are most sensitive to positive incentives; regulatory penalties form rigid constraints on model compliance, and the collaborative benefits play a stable role in the long term. Accordingly, it is necessary to strengthen user feedback incentives, advance regulatory technology empowerment, and optimize institutional collaborative mechanisms. These measures aim to build a governance ecosystem characterized by tripartite collaboration, cost hedging, and risk sharing, thereby providing theoretical support and policy paths for the construction of trustworthy AI and the governance of hallucinations.

Key words: generative artificial intelligence, large language model, AI hallucination, evolutionary game theory, collaborative governance

中图分类号:

TP391.9

闫强,张倩语,魏娜 . 基于演化博弈的生成式人工智能幻觉应对分析[J]. 系统仿真学报, 2026, 38(2): 399-415.

Yan Qiang,Zhang Qianyu,Wei Na . Evolutionary Game-based Analysis of Responses to Hallucinations in Generative Artificial Intelligence[J]. Journal of System Simulation, 2026, 38(2): 399-415.

图/表 12

图1

表1

大模型、用户与监管者的博弈演化收益矩阵

监管者R		积极监管 $z$		消极监管 $1 - z$
用户U		积极反馈 $y$	沉默不语 $1 - y$	积极反馈 $y$	沉默不语 $1 - y$
大模型 M	诚实回答 $x$	$R s h o r t - C s + δ R l o n g + β k$	$R s h o r t - C s$	$R s h o r t - C s + δ R l o n g$	$R s h o r t - C s$
		$η R f - C u + R s o c + δ (W 0 + λ k)$	$δ W 0$	$η R f - C u + R s o c$	0
		$- (1 - θ) C r + δ k + δ R t r u s t$	$- (1 - θ) C r$	0	0
	幻觉生成 $1 - x$	$R s h o r t - δ P l o n g - F$	$R s h o r t$	$R s h o r t - δ P l o n g$	$R s h o r t$
		$η R f - C u + R s o c$	0	$η R f - C u + R s o c$	0
		$- (1 - θ) C r + F$	$- (1 - θ) C r$	0	0

表1

图2

表2

雅可比矩阵对应元素

元素	表达式
$f 11$	$(1 - 2 x) [- C s + z y (β k + F) + y δ (R l o n g + P l o n g)]$
$f 12$	$x (1 - x) [z (β k + F) + δ (R l o n g + P l o n g)]$
$f 13$	$x (1 - x) [y (β k + F)]$
$f 21$	$y (1 - y) (z δ λ k)$
$f 22$	$(1 - 2 y) [η R f - C u + R s o c + x z δ λ k]$
$f 23$	$y (1 - y) (x δ λ k)$
$f 31$	$z (1 - z) [y (δ k + δ R t r u s t - F)]$
$f 32$	$z (1 - z) [x (δ k + δ R t r u s t - F) + F]$
$f 33$	$(1 - 2 z) [- (1 - θ) C r + x y (δ k + δ R t r u s t - F) + y F]$

表2

表3

特征值

均衡点	特征值	稳定条件
$E 1 (0,0, 0)$	$- C s$ $η R f - C u + R s o c$ $- (1 - θ) C r$	$η R f - C u + R s o c < 0$
$E 2 (0,0, 1)$	$- C s + δ (R l o n g + P l o n g) < 0 - C s$ $η R f - C u + R s o c$ $(1 - θ) C r$	不稳定状态
$E 3 (0,1, 0)$	$- C s + δ (R l o n g + P l o n g)$ $- (η R f - C u + R s o c)$ $- (1 - θ) C r + F$	$- C s + δ (R l o n g + P l o n g) < 0$ $- (η R f - C u + R s o c) < 0$ $- (1 - θ) C r + F < 0$
$E 4 (1,0, 0)$	$C s$ $η R f - C u + R s o c$ $- (1 - θ) C r$	不稳定状态
$E 5 (0,1, 1)$	$- C s + β k + F + δ (R l o n g + P l o n g)$ $- (η R f - C u + R s o c)$ $(1 - θ) C r - F$	$- C s + β k + F + δ (R l o n g + P l o n g) < 0$ $- (η R f - C u + R s o c) < 0$ $(1 - θ) C r - F < 0$
$E 6 (1,0, 1)$	$C s$ $η R f - C u + R s o c + δ λ k$ $(1 - θ) C r$	不稳定状态
$E 7 (1,1, 0)$	$- [- C s + δ (R l o n g + P l o n g)]$ $- (η R f - C u + R s o c)$ $- (1 - θ) C r + δ k + δ R t r u s t$	$- C s + δ (R l o n g + P l o n g) > 0$ $η R f - C u + R s o c > 0$ $- (1 - θ) C r + δ k + δ R t r u s t < 0$
$E 8 (1,1, 1)$	$- [- C s + β k + F + δ (R l o n g + P l o n g)]$ $- (η R f - C u + R s o c + δ λ k)$ $- [- (1 - θ) C r + δ k + δ R t r u s t]$	$- C s + β k + F + δ (R l o n g + P l o n g) > 0$ $η R f - C u + R s o c + δ λ k > 0$ $- (1 - θ) C r + δ k + δ R t r u s t > 0$

表3

图3

表4

参数设置

参数	取值	参数	取值	参数	取值
$P l o n g$	10	$C u$	6	$η$	0.6
$R l o n g$	8	$R s o c$	4	$δ$	0.7
$C s$	3	$C r$	4	$θ$	0.6
$F$	5	$R t r u s t$	5	$λ$	0.6
$R f$	4	$k$	5	$β$	0.8

表4

图4

图5

图6

图7

图8

参考文献 35

[1]	Samuel Fosso Wamb, Queiroz Maciel M, Randhawa Krithika, et al. Generative Artificial Intelligence and the Challenges to Adding Value Ethically[J]. Technovation, 2025, 144: 103235.
[2]	Ji Ziwei, Lee N, Frieske R, et al. Survey of Hallucination in Natural Language Generation[J]. ACM Computing Surveys, 2023, 55(12): 248.
[3]	Ouyang Long, Wu J, Jiang Xu, et al. Training Language Models to Follow Instructions with Human Feedback[C]//Proceedings of the 36th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2022: 27730-27744.
[4]	Kalai A T, Nachum O, Vempala S S, et al. Why Language Models Hallucinate[EB/OL]. (2025-09-04) [2025-10-15]. .
[5]	Gintis H. Game Theory Evolving: A Problem-centered Introduction to Modeling Strategic Interaction[M]. 2nd ed. Princeton: Princeton University Press, 2009.
[6]	丁煌, 刘德海. 演化博弈的研究范式、前沿问题和管理启示[J/OL]. 系统工程理论与实践. (2025-10-14) [2025-10-15]. .
	Ding Huang, Liu Dehai. Research Paradigm, Frontier Issues and Management Implications of Evolutionary Game[J/OL]. Systems Engineering-theory & Practice. (2025-10-14) [2025-10-15]. .
[7]	刘泽垣, 王鹏江, 宋晓斌, 等. 大语言模型的幻觉问题研究综述[J]. 软件学报, 2025, 36(3): 1152-1185.
	Liu Zeyuan, Wang Pengjiang, Song Xiaobin, et al. Survey on Hallucinations in Large Language Models[J]. Journal of Software, 2025, 36(3): 1152-1185.
[8]	李自拓, 孙建彬, 陈广州, 等. 大语言模型幻觉检测方法综述[J/OL]. 计算机研究与发展. (2025-10-13) [2025-10-15]. .
	Li Zituo, Sun Jianbin, Chen Guangzhou, et al. Survey of Hallucination Detection for Large Language Models[J/OL]. Journal of Computer Research and Development. (2025-10-13) [2025-10-15]. .
[9]	Zhang Yue, Li Yafu, Cui Leyang, et al. Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models[J]. Computational Linguistics. (2025-09-08) [2025-10-14]. .
[10]	Huang Lei, Yu Weijiang, Ma Weitao, et al. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions[J]. ACM Transactions on Information Systems, 2025, 43(2): 42.
[11]	Chaudhari S, Aggarwal P, Murahari V, et al. RLHF Deciphered: A Critical Analysis of Reinforcement Learning from Human Feedback for LLMs[J]. ACM Computing Surveys, 2025, 58(2): 53.
[12]	Williams M. Multi-objective Reinforcement Learning from AI Feedback[EB/OL]. (2024-06-12) [2025-10-14]. .
[13]	Casper S, Davies X, Shi C, et al. Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback[EB/OL]. (2023-09-11) [2025-10-14]. .
[14]	Song Yan, Cui Mingjia, Wan Fei, et al. AI Hallucination in Crisis Self-rescue Scenarios: The Impact on AI Service Evaluation and the Mitigating Effect of Human Expert Advice[J]. International Journal of Human-Computer Interaction, 2025, 41(22): 14419-14439.
[15]	Wang Zhenhailong, Mao Shaoguang, Wu Wenshan, et al. Unleashing the Emergent Cognitive Synergy in Large Language Models: A Task-solving Agent through Multi-persona Self-collaboration[C]//Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg: ACL, 2024: 257-279.
[16]	Cohen R, Hamri M, Geva M, et al. LM vs LM: Detecting Factual Errors via Cross Examination[EB/OL]. (2023-05-22) [2025-10-15]. .
[17]	Farquhar S, Kossen J, Kuhn L, et al. Detecting Hallucinations in Large Language Models Using Semantic Entropy[J]. Nature, 2024, 630(8017): 625-630.
[18]	Liu Xiaoou, Chen Tiejin, Longchao Da, et al. Uncertainty Quantification and Confidence Calibration in Large Language Models: A Survey[C]//Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. New York: ACM, 2025: 6107-6117.
[19]	张学府. 生成式人工智能服务信息内容安全的三类标准-基于《生成式人工智能服务管理暂行办法》的分析[J]. 中国行政管理, 2024(4): 120-128.
	Zhang Xuefu. Triple Standards for Information Content Security in Generative Artificial Intelligence Services-an Analysis Based on the "Interim Measures for the Administration of Generative Artificial Intelligence Services"[J]. Chinese Public Administration, 2024(4): 120-128.
[20]	吴建南, 马太平, 周磊. 人工智能安全治理的规则体系: 国际比较与协同进路[J]. 中国行政管理, 2024, 40(12): 6-14.
	Wu Jiannan, Ma Taiping, Zhou Lei. The System of Rule for Al Safety/Security Governance: International Comparison and Collaborative Development[J]. Chinese Public Administration, 2024, 40(12): 6-14.
[21]	Taeihagh Araz. Governance of Generative AI[J]. Policy and Society, 2025, 44(1): 1-22.
[22]	Zaidan Esmat, Imad Antoine Ibrahim. AI Governance in a Complex and Rapidly Changing Regulatory Landscape: A Global Perspective[J]. Humanities and Social Sciences Communications, 2024, 11(1): 1121.
[23]	吕悦, 陈旭, 彭子璇, 等. 静态到敏捷: 人工智能监管沙盒治理机制研究[J/OL]. 科学学研究. (2025-09-16) [2025-09-28]. .
	Yue Lü, Chen Xu, Peng Zixuan, et al. From Static to Agile: Governance Mechanism of Artificial Intelligence Regulatory Sandboxes[J/OL]. Studies in Science of Science. (2025-09-16) [2025-09-28]. .
[24]	魏娜, 易兰丽, 张倩语. 技术自主性驱动下的AI政策范式转型研究[J]. 西安财经大学学报, 2025, 38(6): 16-26.
	Wei Na, Yi Lanli, Zhang Qianyu. AI Policy Paradigm Shifts Driven by Technological Autonomy[J]. Journal of Xi'an University of Finance and Economics, 2025, 38(6): 16-26.
[25]	郭海玲, 卫金金, 刘仲山. 生成式人工智能虚假信息协同共治研究[J]. 情报杂志, 2024, 43(9): 121-129, 165.
	Guo Hailing, Wei Jinjin, Liu Zhongshan. Research on Collaborative Governance with Generative Artificial Intelligence in Disinformation[J]. Journal of Intelligence, 2024, 43(9): 121-129, 165.
[26]	林泽源, 余晓, 陈迁, 等. 基于多方演化博弈的人工智能隐私风险协同治理策略研究[J]. 工程管理科技前沿, 2025, 44(3): 25-32.
	Lin Zeyuan, Yu Xiao, Chen Qian, et al. The Multi-evolutionary Game Model of Collaborative Governance Strategies for AI Privacy Risks[J]. Frontiers of Science and Technology of Engineering Management, 2025, 44(3): 25-32.
[27]	Tully S M, Longoni Chiara, Appel G. Lower Artificial Intelligence Literacy Predicts Greater AI Receptivity[J]. Journal of Marketing, 2025, 89(5): 1-20.
[28]	徐浩, 谭德庆, 张敬钦, 等. 群体性突发事件非利益相关者羊群行为的演化博弈分析[J]. 管理评论, 2019, 31(5): 254-266.
	Xu Hao, Tan Deqing, Zhang Jingqin, et al. Evolutionary Game Analysison Herding Behavior of Non-direct Stakeholders in Mass Emergencies[J]. Management Review, 2019, 31(5): 254-266.
[29]	沈映春, 潘淑苓. 大模型产业产学研机制研究-基于演化博弈论和模拟仿真结果[J]. 中国科技论坛, 2024(10): 92-103, 116.
	Shen Yingchun, Pan Shuling. Research on Industry-academia-research Mechanism of LLMs: Based on Evolutionary Game Theory and Simulation Results[J]. Forum on Science and Technology in China, 2024(10): 92-103, 116.
[30]	张健, 张倩语, 廖梦洁, 等. 大宗商品现货交易平台牟利维权"监管破困"策略演化博弈研究[J]. 运筹与管理, 2023, 32(11): 147-154.
	Zhang Jian, Zhang Qianyu, Liao Mengjie, et al. Research on the Evolution Game of the "Regulatory Dilemma" Strategy for Profit-making Rights Protection of Bulk Commodity Spot Trading Platform[J]. Operations Research and Management Science, 2023, 32(11): 147-154.
[31]	汪旭晖, 王佳淏, 仲妍. 智能制造政策助推制造企业数智技术创新——基于供应链视角的动态分析与实证研究[J]. 系统工程理论与实践, 2025, 45(11): 3554-3578.
	Wang Xuhui, Wang Jiahao, Zhong Yan. Intelligent Manufacturing Policy Driving Digital and Intelligent Technology Innovation in Manufacturing Enterprises——A Dynamic Analysis and Empirical Study from the Perspective of Supply Chain[J]. Systems Engineering-Theory & Practice, 2025, 45(11): 3554-3578.
[32]	程乐峰, 杨汝, 王晓刚, 等. 三方多策略式博弈系统的长期演化稳定均衡特性研究[J]. 控制理论与应用, 2021, 38(10): 1631-1661.
	Cheng Lefeng, Yang Ru, Wang Xiaogang, et al. Investigation on Long-term Evolutionarily Stable Equilibrium Characteristics of Three-party Multi-strategy Game Systems[J]. Control Theory & Applications, 2021, 38(10): 1631-1661.
[33]	Friedman D. Evolutionary Economics Goes Mainstream: A Review of the Theory of Learning in Games[J]. Journal of Evolutionary Economics, 1998, 8(4): 423-432.
[34]	Weibull J W. Evolutionary Game Theory[M]. Cambridge: MIT Press, 1995.
[35]	范文慧, 蒋沅. 人工智能时代的仿真科学与工程思考[J]. 系统仿真学报, 2025, 37(7): 1607-1623.
	Fan Wenhui, Jiang Yuan. Thinking on Simulation Science and Engineering in the Era of Artificial Intelligence[J]. Journal of System Simulation, 2025, 37(7): 1607-1623.