用 Jev 做 agent 每步推理强度的路由控制

vechen @miu21590原推发布于 2026.09.22

在 X 上打开原推

Inspired by Ares paper (https://arxiv.org/abs/2603.07915), which used a fine-tuned Qwen3-1.7B to pick reasoning effort before every agent step. Same idea, but Jev is a much better fit for this control loop: fast enough that the routing overhead is basically invisible. It’s especially practical now that Astra and the latest Claude models can change effort mid-conversation without invalidating the prompt cache.