Google 提出自改进智能体框架 GUARD:冻结基座模型,只改提示词与记忆
做 Agent 的开发者应该看看这篇:Google 把自改进 harness 的老毛病(过拟合、遗忘、记忆污染)拆解得很清楚。GUARD 靠冻结模型、只改提示词和记忆来实现改进,比 Meta-Harness 少用一半 token 还多拿 6 分,适合需要长期自演化 Agent 的团队参考。
Google 在新论文(arXiv 2611.23476)中提出 GUARD,一种自改进智能体 harness 方法。它保持基座模型冻结,让智能体自行编辑提示词、技能、工具规则和记忆,只有当某次编辑带来的收益超过其破坏的行为时才被接受,以此避免局部过拟合、灾难性遗忘和记忆污染。
实验数据上,GUARD 在 43 个 hosted SDK 任务上达到 80.80% 准确率,比起始 harness 高 8.89 个百分点(p = 0.008);比 Meta-Harness 高 6.28 个百分点,且每任务 token 消耗减少 54.5%。在程序性记忆任务上,更新正确率达 90.1%,且零记忆污染,而激进改写策略虽然分数相近但污染翻了三倍。
机制上,GUARD 把每个任务拆成执行前的自检项,每轮用它们评估轨迹,锁定智能体已掌握的规则,并优先修复新被破坏的行为。测试时单个冻结 harness 一次前向运行即可完成。
Banger paper from Google on self-improving agent harnesses. (bookmark it) It addresses common issues in self-improving agent harnesses like local overfitting, catastrophic forgetting, and memory pollution. They propose GUARD, which keeps the base model frozen and lets the agent edit its own prompts, skills, tool rules, and memory. An edit is accepted only when its gains outweigh the behaviors it breaks. - On 43 hosted SDK tasks, GUARD reached 80.80% accuracy, 8.89 points over the starting harness (p = 0.008). - It beat Meta-Harness by 6.28 points and used 54.5% fewer tokens per task. - On procedural memory tasks, it led update correctness at 90.1% with zero memory pollution. Aggressive rewriters scored similarly and tripled pollution. How does it know what broke? Each task is split into self-checks before execution. Every round grades trajectories against them, pins rules the agent already handles, and ranks newly broken behaviors first. At test time, one frozen harness runs in a single pass. arxiv.org/abs/2611.23476 Chat with Paper: academy.dair.ai/papers/guard-s…📊 14 ⚡
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」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:GUARD、智能体、Google、论文,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「elvis」(@omarsar0),经 AITOP 多维评分管线结构化提炼。