📄 中文摘要
SkillsBench 是一项雄心勃勃的研究，涉及 84 个任务、11 个领域、7 个编码代理和 7,308 条轨迹。该研究评估了在三种条件下的任务表现：无技能、经过策划的技能和自生成的技能。虽然研究结果显示，经过策划的技能可以提高编码代理的表现，但在深入分析其方法论后，发现了一些潜在的问题。这些问题可能影响结果的可靠性和可重复性，从而使得研究的实际应用价值受到质疑。
我为何不会在 SkillsBench 上采取行动
出处: Why I Wouldn't Act on SkillsBench
发布: 2026年2月25日
📄 English Summary
Why I Wouldn't Act on SkillsBench
SkillsBench is an ambitious study involving 84 tasks, 11 domains, 7 coding agents, and 7,308 trajectories. It evaluates task performance under three conditions: no skills, curated skills, and self-generated skills. While the results indicate that curated skills can enhance the performance of coding agents, a deeper analysis of the methodology reveals potential issues. These issues may affect the reliability and reproducibility of the results, thereby questioning the practical applicability of the research findings.