Google 发布 EfficientSkill:先学准再剪枝,让智能体技能更省 token
做智能体的都遇到过技能文件越写越长、token 成本失控的问题。Google 这篇给出可复制的两阶段流程:先把准确率练上去,再剪枝瘦身。四个模型、三个基准验证,成本最多降 67%,准确率还涨了 14 分以上,自己搭 agent 技能库的话直接照搬流程就行。
Google 发布 EfficientSkill 论文,针对智能体技能文件随失败不断膨胀的问题提出两阶段优化方法。传统技能优化器每遇一次失败就追加一条规则,导致每轮读取的 token 越来越多,长循环容易超时。
EfficientSkill 的第一阶段在无长度惩罚条件下修复失败,直到技能达到目标准确率;第二阶段再修剪指令、把可复用代码移入 workspace 工具库,同时要求准确率保持在目标附近。作者强调顺序很重要:从空种子开始剪枝会停滞。
实验覆盖四个模型与 MATH-500、ALFWorld、SpreadsheetBench 三个基准,12 个配置中有 11 个输出 token 下降,最高节省 80%。准确率在所有基准上都提升,SpreadsheetBench 上 Gemini 3.8 Flash 提升 14.3 分,Grok 4.1 Fast 提升 17.5 分,服务成本下降 25% 到 67%。
Great practical paper from Google on building efficient agent skills.
For cheaper agents, learn a skill for accuracy first, then prune it for length. This has also worked well in our own experiments.
Skill optimizers usually add a new rule after every failure. The skill grows, the agent reads more tokens every turn, and long loops start to time out.
EfficientSkill runs two phases.
The first repairs failures with no length penalty until the skill reaches a target accuracy. The second prunes instructions and moves reusable code into workspace utilities, while accuracy has to stay close to that target. Pruning from the empty seed stalls, so the order matters.
Across four models on MATH-500, ALFWorld, and SpreadsheetBench, output tokens fall on 11 of 12 configurations, up to 80% of the time. Accuracy also rises on every benchmark, including +14.3 points with Gemini 3.8 Flash and +17.5 points with Grok 4.1 Fast on SpreadsheetBench. Serving cost drops 25% to 67%.
https://t.co/W1XsCgoKV2
Chat with Paper: https://t.co/ml7ZignKV7
本条情报归属 AITOP「论文」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」「中美大模型技术代差与商业化对照」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:EfficientSkill、Google、Gemini、智能体,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明:本文原始线索采编自「DAIR.AI」(@dair_ai),经 AITOP 多维评分管线结构化提炼。