DeepMind 这篇论文告诉你，别被路由器的漂亮数字骗了——模型都答一样或乱分配，路由就是摆设。
DeepMind 新论文提出 LLM 路由器的两个关键属性：行为分化（模型间输出不同）和表面形式重写稳定性（同一查询的不同表述应路由到同一专家）。实验表明，仅看准确率和成本可能掩盖无意义的路由（如冗余模型池或分配不一致）。论文编号 arxiv.org/abs/2607.09197，提供实用检查方法。
Great paper from DeepMind on effective model routing strategies.
Great paper from DeepMind on effective model routing strategies. DAIR.AI @dair_ai New research from Google DeepMind on effective model routing. LLM routers get judged on accuracy and cost. Both can look great while the router is meaningless. If every model in your society responds the same way, routing is vacuous, you get the same answers no matter where queries land. And if paraphrases of one query get sent to different experts, the router is unstable, so its assignments carry no real signal. The argument is that two properties decide whether routing means anything. Namely, behavioural differentiation of the actors and stability under surface-form rewrites. And they both are orthogonal to task accuracy. What's the practical take here? If you use mixture-of-agents or model routing, your overall accuracy can hide a router that operates over a redundant society or assigns queries inconsistently. These two checks catch the routers that look good and do nothing. Paper: arxiv.org/abs/2607.09197 Learn to build effective AI agents in our academy: academy.dair.ai 🔗 View Quoted Tweet 💬 3 🔄 0 ❤️ 17 👀 3444 📊 6 ⚡