AI 中文总结
研究探讨市场中人工智能-智能体对齐，提出将基础市场设计视为该对齐层次，通过对市场核心形式化建模，利用理论计算机科学严谨性构建透明盒模型，支持激励分析与机制设计，使期望行为受青睐，不良行为难维持。
AI 中文摘要
本文认为，市场中的人工智能-智能体对齐不应仅被理解为智能体的属性，还应被视为智能体所处交互基础设施的属性。在金融市场中，该基础设施即市场核心，是决定订单如何进入、交互、匹配、持续和稳定的规则系统。若此基础交互层允许或奖励不良行为，那么智能体的高层对齐可能不足。我们提议将基础市场设计视为人工智能-智能体对齐的一个层次。除了计算经济学在对智能体、策略和学习进行建模方面的重要工作外，我们关注一个互补但更基础的层次：市场核心本身的形式化建模。市场设计，尤其是在核心机制层面，可受益于理论计算机科学的严谨性。这给出了一个市场的透明盒模型，其核心属性可被正式指定和推理。它还让我们将交易场所视为一个将常驻订单与流入订单流相结合的计算过程，并询问哪种计算模型（可能尚不清楚）自然地处于其核心。这种观点对于由自适应或人工智能智能体组成的市场尤为相关。此类智能体可能会了解机制奖励的内容，包括速度、延迟、流动性提供或操纵。这些行为不仅是单个智能体的属性，还可能从智能体-机制系统中出现。因此，我们认为市场核心的透明形式模型可以支持面向激励的分析和机制设计，在这种机制中，期望的行为在结构上受到青睐，而不良行为则更难维持。
英文摘要
This paper argues that AI-agent alignment in markets should not be understood only as a property of agents, but also as a property of the interaction infrastructure in which agents act. In financial markets, this infrastructure is the market core: the rule system that determines how orders enter, interact, match, persist, and stabilize. If this fundamental interaction layer allows or rewards undesired behaviour, then higher-level alignment of agents may be insufficient. We propose to view fundamental market design as a layer of AI-agent alignment. Alongside the important work of computational economics in modelling agents, strategies, and learning, we focus on a complementary but more fundamental layer: the formal modelling of the market core itself. Market design, especially at the level of the core mechanism, can benefit from a rigour characteristic of theoretical computer science. This gives a transparent-box model of the market, whose core properties can be formally specified and reasoned about. It also lets us treat the trading venue not as a static order book, but as a computational process combining resident orders with incoming order flow, and ask which computational model, perhaps yet unknown, naturally lies at its core. This perspective is especially relevant for markets populated by adaptive or AI agents. Such agents may learn what the mechanism rewards, including speed, delay, liquidity provision, or manipulation. These behaviours are not only properties of individual agents, but may emerge from the agent-mechanism system. We therefore argue that transparent formal models of market cores can support incentive-oriented analysis and the design of mechanisms in which desirable behaviours are structurally favoured and undesirable behaviours are harder to sustain.
CommentsAccepted as at the EC'26 Workshop on Incentive-Based AI Alignment, co-located with the 27th ACM Conference on Economics and Computation, Rome, Italy, July 2026. This version is prepared for public dissemination following workshop acceptance
