Agent Arena 榜单将 Bash Recovery 指标升级为 Bash Reliability
Agent Arena 把 bash 失败判定改得更细了，只算模型自己的锅，用得干净的还能加分，选编程智能体时可以参考。
Agent Arena 榜单将原来的 Bash Recovery 指标替换为 Bash Reliability。新指标会读取每条 bash 调用的目的、结果和错误原因，判断错误是否在同一次调用内修复，只统计智能体自身造成的失败。命令行使用干净的模型现在会因一次到位获得加分。回填后的分数将陆续更新到公开的 Hugging Face 数据集。
We evolved Bash Recovery to Bash Reliability for the Agent Arena leaderboard. - A sharper read of what counts as a failure: we now read each bash call’s purpose, outcome and error cause, and whether the error was fixed within the call. The new signal now captures only mistakes the agent actually has made. - Crediting reliable bash use: Bash Reliability measures agent-caused failures per non-exempt call, across every session. Models with a clean command-line use now get credit for getting it right the first time. Backfilled scores are coming to our public Hugging Face datasets soon. Learn more about Agent Arena signal definitions and Bash Reliability at: arena.ai/resources/agen… 💬 1 🔄 0 ❤️ 6 👀 2213 📊 2 ⚡
