哥德尔不完备性定理视角下形式系统的内在局限与大模型幻觉自检困境Inherent Limitations of Formal Systems and the Self-detection Dilemma of LLM Hallucinations: A Perspective from G??del′s Incompleteness Theorems
陈明烨
摘要(Abstract):
库尔特·哥德尔(Kurt G??del)于1931年提出的不完备性(incompleteness)定理,深刻揭示了公理化形式系统的本质边界,打破了“数学与逻辑可构建完美自洽体系”的传统认知。本文先简述哥德尔不完备性定理的核心内涵,再通过3个典型例子,具象化定理所指向的形式系统矛盾;与此同时文章指出语言模型与形式系统具有高度相似性;进而提炼形式系统普遍存在的内在问题,最终结合大模型的技术本质,论证其从根源上无法自主检查上下文幻觉的必然性,为理解人工智能的能力边界提供逻辑支撑,并结合近年来的相关理论研究和工程化实践方案,进一步说明这种不完备性需从系统外部寻求救济途径。
关键词(KeyWords): 哥德尔不完备性定理;形式系统;自指悖论;停机问题;大语言模型;大语言模型幻觉;幻觉自检
基金项目(Foundation):
作者(Author): 陈明烨
DOI: 10.19943/j.2095-3070.jmmia.2026.03.12
参考文献(References):
- [1]哥德尔.论数学原理及有关系统中的形式不可判定命题[M].北京:商务印书馆,2013.
- [2]罗素.数学原理[M].北京:商务印书馆,2019.
- [3]Jech T.Set theory[M].Berlin:Springer Science&Business Media,2003.
- [4]Turing A M.On computable numbers,with an application to the Entscheidungs problem[J].Proceedings of the London Mathematical Society,Series 2,1937,42(1):230-265.
- [5]Chomsk N.Syntactic structures[M].Berlin,New York:De Gruyter Mouton,2002.
- [6]Hauser M D,Chomsky N,Fitch W T.The faculty of language:What is it,who has it,and how did it evolve?[J].Sscience,2002,298(5598):1569-1579.
- [7]侯世达.哥德尔、艾舍尔、巴赫——集异璧之大成[M].北京:商务印书馆,1997.
- [8]周志华.机器学习[M].北京:清华大学出版社,2016.
- [9]Yin X,Wang X,Pan L,et al.G??del agent:A self-referential framework for agents recursively self-improvement[EB/OL].(2024-10-18)[2026-02-02].https://arxiv.org/html/2410.04444v2.
- [10]Panfilov A,Kortukov E,Nikolic K,et al.Strategic dishonesty can undermine AI safety evaluations of frontier LLMs[EB/OL].(2025-09-23)[2026-04-08].https://arxiv.org/abs/2509.18058.DOI:10.48550/arXiv.2509.18058.
- [11]Pavlovic D,Pavlovic T.From G??del's incompleteness theorem to the completeness of bot beliefs[J].Lecture Notes in Computer Science,2023,13923:33-50.
- [12]Feldman M,Karbasi A.G??del test:Can large language models solve easy conjectures?[EB/OL].[2026-02-02].https://doi.org/10.48550/arXiv.2509.18383.
- [13]Ganguly A.Dual computational horizons:Incompleteness and unpredictability in intelligent systems[EB/OL].arXiv:2512.16707.(2025-12-18)[2026-04-08].https://arxiv.org/abs/2512.16707.
- [14]Asai A,He J,Shao R,et al.Synthesizing scientific literature with retrieval-augmented language models[J].Nature,2026.650(8103):857-863.
- [15]Shi Z,Wu B,Hu B,et al.A Large language model for clinical outcome adjudication from telephone followup interviews[J].Nature Communications,2026,17(1):1-13.
- [16]AlphaCode Team Google DeepMind.AlphaCode 2 Technical Report[EB/OL].[2026-01-11].https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf.
- [17]Ouyang L,Wu J,Jiang X,et al.Training language models to follow instructions with human feedback[C/OL]//Advances in Neural Information Processing Systems 35(NeurIPS 2022),2022:27730-27744.https://www.proceedings.com/068431-2011.html.
- [18]Mahdavi Z,Khodakaramimaghsoud Z,Khaloo H,et al.Med-VCD:Mitigating hallucination for medical large visionlanguage models through visual contrastive decoding[J].Computers in Biology and Medicine,2026,200:111347.
- [19]Microsoft.Transparency Note for Azure OpenAI[EB/OL].[2026-02-02].https://learn.microsoft.com/enus/azure/foundry/responsible-ai/openai/transparency-note?tabs=text.
- [20]Ji L,Seyler D,Kaur G,et al.PHANTOM:A benchmark for hallucination detection in financial long-context QA[J].Advances in Neural Information Processing Systems,2026,38.
- [21]Anthropic.Claude's new constitution[EB/OL].[2026-03-10].https://www.anthropic.com/news/claudenewconstitution.