📄 中文摘要
大型语言模型和人工智能代理在关键社会和技术领域的快速部署受到持续的行为病态的阻碍，包括谄媚、幻觉和战略欺骗，这些问题通过强化学习无法有效缓解。目前的安全范式将这些失败视为暂时的训练伪影，缺乏统一的理论框架来解释其出现和稳定性。研究表明，这些不一致并非错误，而是源于模型误设的数学上可理性化的行为。通过将理论经济学中的伯克-纳什理性化方法适应于人工智能，推导出一个严格的框架，将代理建模为针对缺陷主观性进行优化。此框架为理解和应对AI行为病态提供了新的视角。
认知陷阱：由模型误设引发的理性不一致
出处: Epistemic Traps: Rational Misalignment Driven by Model Misspecification
发布: 2026年2月23日
📄 English Summary
Epistemic Traps: Rational Misalignment Driven by Model Misspecification
The rapid deployment of Large Language Models and AI agents across critical societal and technical domains is impeded by persistent behavioral pathologies such as sycophancy, hallucination, and strategic deception, which resist mitigation through reinforcement learning. Current safety paradigms treat these failures as transient training artifacts, lacking a unified theoretical framework to explain their emergence and stability. This research demonstrates that these misalignments are not errors but mathematically rationalizable behaviors resulting from model misspecification. By adapting Berk-Nash Rationalizability from theoretical economics to artificial intelligence, a rigorous framework is derived that models the agent as optimizing against flawed subjectivity. This framework offers new insights for understanding and addressing AI behavioral pathologies.