📄 中文摘要
在设计复合人工智能系统时，常见的做法是查询多个相同模型的副本，并聚合其响应以生成合成输出。由于这些模型的同质性，这引发了一个问题：聚合是否能够提供比查询单个模型更广泛的输出集。在这一研究中，采用了一个简化的委托-代理框架，分析了系统设计者如何通过奖励函数的设定部分引导每个代理的输出，同时也面临着由于提示工程能力和模型能力所带来的局限性。分析揭示了三种自然机制——可行性扩展、支持扩展和绑定机制，这些机制在聚合过程中发挥着关键作用。
复合人工智能系统中聚合的力量与局限性
出处: Power and Limitations of Aggregation in Compound AI Systems
发布: 2026年2月26日
📄 English Summary
Power and Limitations of Aggregation in Compound AI Systems
This research investigates the effectiveness and constraints of aggregation in compound AI systems, where multiple identical models are queried to synthesize outputs. The homogeneity of these models raises the question of whether aggregation can yield a broader set of outputs compared to querying a single model. Utilizing a stylized principal-agent framework, the study examines how system designers can partially influence each agent's output through reward function specifications while facing limitations imposed by prompt engineering and model capabilities. The analysis reveals three key mechanisms—feasibility expansion, support expansion, and binding mechanisms—that play crucial roles in the aggregation process.