Google DeepMind提出Pandora路由框架优化模型选择成本
Google DeepMind’s new routing idea is trying to solve a great practical question…
这篇论文提出了极具实用价值的模型路由优化思路，用经济学视角解决算力分配痛点，数据详实且对比鲜明，值得关注系统架构优化的同学深入阅读。
Google DeepMind’s new routing idea is trying to solve a great practical question.
Google DeepMind 的新路由思路试图解决一个非常实际的问题。
Routing is supposed to save compute, but deciding where to route can itself cost compute.
路由旨在节省计算资源，但决定将请求路由到何处本身也会消耗计算资源。
Paper says spend more on model selection only when the better decision is worth the extra cost.
论文指出，只有当更优的决策带来的收益超过额外成本时，才应在模型选择上投入更多。
Google DeepMind frames that decision as a Pandora’s Box problem.
Google DeepMind 将该决策框架化为“潘多拉魔盒”问题。
Every specialist gets a cheap, noisy score first. The router pays for a stronger estimate only when the expected value of that extra information is higher than its cost.
每个专家模型首先会获得一个廉价但有噪声的评分。仅当额外信息带来的预期价值高于其成本时，路由器才会付费获取更精确的估计值。
That changes the economics of routing.
这改变了路由的经济模型。
Across MATH, RAG, and EmbedLLM, Pandora’s Router had the lowest or tied-lowest combined routing regret and inspection cost on average across the tested cost settings.
在 MATH、RAG 和 EmbedLLM 任务中，Pandora’s Router 在测试的各种成本设置下，平均综合路由后悔值和检查成本最低或并列最低。
On EmbedLLM, with more than 100 routing targets, its average inspection cost was 0.075 versus 1.986 for always running the expensive estimator, while routing regret also fell from 0.370 to 0.311.
在 EmbedLLM 上，面对超过 100 个路由目标，其平均检查成本为 0.075，而始终运行昂贵估计器的成本为 1.986；同时，路由后悔值也从 0.370 降至 0.311。
– arxiv. org/abs/2608.20316
– arxiv.org/abs/2608.20316
Title: "Pandora's AI Model Routing Box: Efficient Allocation with Costly Value Estimation"
标题：《潘多拉的 AI 模型路由盒：具有高昂价值估计成本的效率分配》
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力