Google DeepMind 研究 AI 智能体竞价市场中的拍卖效率
DeepMind 这篇论文把模型路由问题变成拍卖机制设计，还给出 1-1/e² 这个可验证的效率界，做智能体路由的人应该看看。
Google DeepMind 的论文研究多个 AI 智能体通过竞价来回答用户提示词的市场机制，每个智能体的出价等于预期收益减去推理成本。由于一价拍卖会让投标者压低出价，提示词可能流向价值并非最高的智能体。论文证明两种条件可以减小损失：当多个智能体对同一提示词估值接近时（例如通过蒸馏得到便宜模型），竞争会限制损失；当估值主要取决于提示词难度时，效率可达最优的 1 - 1/e^2，高于一般情形的 1 - 1/e 下界。在 EmbedLLM 路由数据集上的模拟结果与理论一致。
Fascinating paper from Google DeepMind.
They studied a market where AI agents bid to answer each user prompt.
Each agent bids its expected reward minus its inference cost. A first-price auction is simple to run, but bidders shade their bids, so a prompt can go to an agent that would not have produced the most value.
The paper shows two features of agent markets that keep this loss small. When several agents value a prompt close to the best one, for example because strong models are distilled into cheaper ones, competition limits the loss. When values mostly depend on how hard the prompt is, efficiency approaches 1 - 1/e^2 of optimal, well above the general bound of 1 - 1/e.
They find that simulations on the EmbedLLM routing dataset match the theory.
https://t.co/7cLDnMozvy
Chat with Paper: https://t.co/wXBmyb3qBF
