论文：Jev 接管 agent 决策，前沿模型调用减少 72.7%

Chrome @0xchromium原推发布于 2026.09.27

在 X 上打开原推

two researchers just published a paper where Jev runs an AI agent's decisions and a frontier LLM only steps in when it's truly needed the twist: the expensive model stopped being the brain and became the specialist here's what the paper shows: > Jev takes every bounded decision in the agent loop, anything with a fixed set of answers > the LLM gets called only for writing, or for the rare decision Jev flags as uncertain > on a frozen 100-task benchmark, the agent finished 95 of the 100 tasks > it made 72.7% fewer calls to the expensive model than an agent running on that model alone > with Qwen, Kimi or DeepSeek as the fallback, the cut stayed between 66% and 72% the result: you stop paying frontier prices for decisions a typed model handles on its own how to write the questions that make Jev's confidence worth trusting is what my article covers

这条是引用着另一条写的：x.com/1450821446859624454/status/210275561781110…