ReJev 用 MiniCPM5-2B 做智能体决策，准确率升至 80.50%
Jev has sparked a compelling question: why generate a paragraph when your agent just needs to make a decision?
独立社区项目 ReJev 通过 LoRA 后训练，将面壁智能开源 2B 模型 MiniCPM5-2B 适配为「状态+问题+候选选项→一个决策」的任务，在 1,892 条密封留出样本上准确率从 51.11% 升至 80.50%（+29.39 个百分点），报告评测中无效输出为 0%。该项目累计 Modal 账单约 5.31 美元，作者称这是早期任务特定结果，并非与 Jev 持平，但可用于探索智能体路由与工作流控制。
正文
当前显示原文，中文全文译文尚未就绪。
Jev has sparked a compelling question: why generate a paragraph when your agent just needs to make a decision?
Choosing a route, classifying an input, selecting the next action—many steps in an AI workflow need a clear choice from known options.
That’s the idea behind ReJev, an independent community project exploring Jev-style decision-making with MiniCPM5-2B. Through LoRA post-training, it adapts our open 2B model to a focused task: State + question + candidate options → one decision. On the project’s sealed holdout of 1,892 samples: 📈 Accuracy rose from 51.11% to 80.50% (+29.39 percentage points) 🎯 0% invalid outputs in the reported evaluation 💰 ~$5.31 in cumulative Modal app billing, including earlier experimental overhead
What makes this interesting goes beyond the accuracy gain: it gives developers a concrete experiment in teaching a small, open model to make bounded decisions—the kind of capability worth exploring for agent routing and workflow control.
This is an early, task-specific result, rather than evidence of parity with Jev. But it opens up a practical question for builders:
Which decisions in your agent stack could a specialized 2B model handle? 🔗 Explore ReJev：https://t.co/iIUxmiJIYa 🤗 Build with MiniCPM5-2B：https://t.co/FZOMTZhBjq