NVIDIA 新论文提出 ACES 评估框架，指出技能文档得分高不代表智能体运行时真能受益——技能主要贡献是发现与工作流顺序，而非最终答案质量。ACES 固定任务、模型、沙箱等变量，仅对比有无目标技能两次运行的奖励差，定义为 Skill Lift。在生产技能实测中，结构与裁判评分与 Skill Lift 相关性仅 −0.0181 和 −0.0266，几乎为零。
New Nvidia paper.
A high-scoring skill document is not evidence that the skill helps the agent at runtime.
Most of what an agent skill adds is discovery and workflow order, not a better final answer. So grade the trajectory, since final-answer scoring hides whether the agent ever found or followed the skill.
So run the same task twice, once with the skill loaded and once without, then compare.
NVIDIA's ACES holds task, model, harness, sandbox, grader, and supporting skills fixed, varies only whether the target skill is present, and reports the paired reward difference as Skill Lift.
On production skills with both a document score and a live run, structural and judge scores correlate with measured lift at −0.0181 and −0.0266, indistinguishable from zero.
– arxiv. org/abs/2608.20614
Title: "Evaluating Skills, Not Just Agents: Agentic Continuous Evaluation of Skills"
来源：Rohan Paul · x.com