论文73°
加提示词后同一模型失败率从 75.1% 降至 6.3%
With hints, the unchanged model avoided the original failure in 93.7% of cases, up from 75.1% In a ...
精选理由
Perplexity 放出实测数据：不换模型只加提示词，失败率就从 75.1% 降到 6.3%，做工具调用的值得对比下自己的数字。
Perplexity 公布了一组测试数据：不给模型做任何更新，仅在推理时加入提示词（hints），模型在原本会失败的案例中 93.7% 的情况下成功避开失败，此前为 75.1%。另一项独立实测中，两版模型在不加推理时提示的条件下，工具调用失败率从 2.24% 降到 1.77%。数据说明提示词设计与模型训练都能降低工具调用错误，两条路径可分别量化。
原文 · Perplexity
With hints, the unchanged model avoided the original failure in 93.7% of cases, up from 75.1% In a ...
With hints, the unchanged model avoided the original failure in 93.7% of cases, up from 75.1% In a separate live test, tool-call failures fell from 2.24% to 1.77% between trained versions, without hints at inference time. 💬 1 🔄 0 ❤️ 3 👀 2031 📊 1 ⚡
