yb2698· @yb2698 · X·· 5 小时前AI 评分
Pi harness 优化 Qwen 模型效果不佳
为 Qwen 模型优化 Pi harness？👀 我们尝试通过调整 harness 的各个组件来优化 Pi harness 以适配 Qwen 模型系列。尽管在 ALE（Agents Last Exam）上有初步发现，但似乎在 Terminal-Bench Lite 上没什么提升，这是为什么？ 实验设置和发现见线程 🧵
Pi harness optimized for Qwen models? 👀
We tried to optimize the Pi harness for Qwen model series by trying to tweaking various components from the harness. Despite early findings on ALE (Agents Last Exam), it seems nothing much improves with Terminal-Bench Lite, why is that?
Experiment setup and findings in the thread 🧵
来源：yb2698 · x.com