{"version":"1.0","generated_at":"2026-10-05T10:46:22.188489","id":3234,"slug":"google-researchers-find-a-way-to-keep-self-improving-ai-agents-from-memorizing-their-tests","title":"谷歌研究人员找到防止自我改进AI智能体死记测试的方法","summary":"Google Cloud AI Research与多所大学提出RRSI，通过约束递归自我改进流程，防止AI代理在测试任务上过拟合，同时在未见基准上提升表现并控制计算成本。","abstract":"谷歌研究人员找到防止自我改进AI智能体死记测试的方法 Google Cloud AI Research与多所大学提出RRSI，通过约束递归自我改进流程，防止AI代理在测试任务上过拟合，同时在未见基准上提升表现并控制计算成本。 Google Cloud AI Research与多所大学提出RRSI，通过约束递归自我改进流程，防止AI代理在测试任务上过拟合，同时在未见基准上提升表现并控制计算成本。 原贴提到：AI agents that keep optimizing their own working environment quickly ten 来源：the-decoder.com 持续优化自身工作环境的AI代理，很快就会过度专门化于它们的测试任务。来自Google Cloud AI Research和几所大学的一种新方法，旨在防止这种情况，同时降低计算成本。 现代AI代理把固定的语言模型包裹在所谓的harness中，这是一个由提示、工作流、工具、记忆和逻辑组成的框架，控制模型在每一步看到什么。harness决定代理在修改文件前是否读取了正确的文件，是否能从错误中恢复，以及是否能干净地交付结果。根据一篇新研究论文，近来代理的许多进展来自对harness的工作，而不是来自新模型。 直到最近，这还是手工完成的。人们审查失败的运行，并手动修补harness。较新的方法通过让一个语言模型根据测试任务的反馈，反复重写harness本身，来自动化这一循环。研究人员称这是一种实用形式的递归自我改进。系统产生反馈，用它来优化harness，而harness反过来控制系统自身的行为。 论文表…","access_level":"member","access_label":"免费阅读","access_mode":"full","is_preview":false,"canonical_url":"https://opc.beizhux.com/content/3234/google-researchers-find-a-way-to-keep-self-improving-ai-agents-from-memorizing-their-tests","html_url":"https://opc.beizhux.com/content/3234/google-researchers-find-a-way-to-keep-self-improving-ai-agents-from-memorizing-their-tests","json_url":"https://opc.beizhux.com/content/3234/google-researchers-find-a-way-to-keep-self-improving-ai-agents-from-memorizing-their-tests.json","published_at":"2026-10-04T20:40:37","updated_at":"2026-10-05T10:08:38","category":{"slug":"ai-business","name":"AI小生意项目库"},"source":{"site":"the-decoder.com","author":"Jonathan Kemper","url":"https://the-decoder.com/google-researchers-find-a-way-to-keep-self-improving-ai-agents-from-memorizing-their-tests/"},"tags":["Agent Harness","AI Agent","AI代理","Google","Google Cloud AI Research","Harness","RRSI","基准测试","泛化能力","递归自我改进"],"topics":[{"slug":"agent-workflow","name":"Agent工作流","url":"https://opc.beizhux.com/topics/agent-workflow","reason":"这篇内容命中「Agent、智能体」等主题信号。"},{"slug":"ai-side-business","name":"AI副业","url":"https://opc.beizhux.com/topics/ai-side-business","reason":"这篇内容命中「项目、小生意、变现」等主题信号。"}],"keywords":["AI小生意项目库","Agent工作流","AI副业","工具","自动化","模型","Cursor","Agent","智能体","工作流","Claude Code","AI编程"],"questions":[{"question":"谷歌研究人员找到防止自我改进AI智能体死记测试的方法主要讲什么？","answer":"Google Cloud AI Research与多所大学提出RRSI，通过约束递归自我改进流程，防止AI代理在测试任务上过拟合，同时在未见基准上提升表现并控制计算成本。"},{"question":"这篇文章最值得关注的要点是什么？","answer":"Google Cloud AI Research与多所大学提出RRSI，通过约束递归自我改进流程，防止AI代理在测试任务上过拟合，同时在未见基准上提升表现并控制计算成本。；原贴提到：AI agents that keep optimizing their own working environment quickly ten；来源：the-decoder.com"},{"question":"这篇文章和哪些AI专题相关？","answer":"它适合放在Agent工作流、AI副业专题里阅读。 关联原因：这篇内容命中「Agent、智能体」等主题信号。；这篇内容命中「项目、小生意、变现」等主题信号。"},{"question":"阅读这篇文章建议先理解哪些关键词？","answer":"建议先理解工具、自动化、模型、Cursor、Agent这些关键词，再结合正文判断工具、机会或风险是否值得进入自己的工作流。"}],"terms":[{"slug":"tools","name":"工具","definition":"在AI觉醒星球里，「工具」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/tools","path":"/glossary/tools","json_url":"https://opc.beizhux.com/glossary/tools.json"},{"slug":"automation","name":"自动化","definition":"在AI觉醒星球里，「自动化」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/automation","path":"/glossary/automation","json_url":"https://opc.beizhux.com/glossary/automation.json"},{"slug":"ai-model","name":"模型","definition":"在AI觉醒星球里，「模型」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/ai-model","path":"/glossary/ai-model","json_url":"https://opc.beizhux.com/glossary/ai-model.json"},{"slug":"cursor","name":"Cursor","definition":"在AI觉醒星球里，「Cursor」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/cursor","path":"/glossary/cursor","json_url":"https://opc.beizhux.com/glossary/cursor.json"},{"slug":"agent","name":"Agent","definition":"在AI觉醒星球里，「Agent」属于「Agent工作流」方向。聚合Agent、AI编程、Claude Code、Cursor、工作流自动化等内容，帮助你把智能体能力落到真实任务链路。 把智能体当成任务系统，而不是聊天窗口。","topic_slug":"agent-workflow","topic_name":"Agent工作流","topic_title":"Agent工作流：智能体、AI编程与自动化协作","topic_url":"https://opc.beizhux.com/topics/agent-workflow","topic_path":"/topics/agent-workflow","url":"https://opc.beizhux.com/glossary/agent","path":"/glossary/agent","json_url":"https://opc.beizhux.com/glossary/agent.json"},{"slug":"ai-agent","name":"智能体","definition":"在AI觉醒星球里，「智能体」属于「Agent工作流」方向。聚合Agent、AI编程、Claude Code、Cursor、工作流自动化等内容，帮助你把智能体能力落到真实任务链路。 把智能体当成任务系统，而不是聊天窗口。","topic_slug":"agent-workflow","topic_name":"Agent工作流","topic_title":"Agent工作流：智能体、AI编程与自动化协作","topic_url":"https://opc.beizhux.com/topics/agent-workflow","topic_path":"/topics/agent-workflow","url":"https://opc.beizhux.com/glossary/ai-agent","path":"/glossary/ai-agent","json_url":"https://opc.beizhux.com/glossary/ai-agent.json"},{"slug":"workflow","name":"工作流","definition":"在AI觉醒星球里，「工作流」属于「Agent工作流」方向。聚合Agent、AI编程、Claude Code、Cursor、工作流自动化等内容，帮助你把智能体能力落到真实任务链路。 把智能体当成任务系统，而不是聊天窗口。","topic_slug":"agent-workflow","topic_name":"Agent工作流","topic_title":"Agent工作流：智能体、AI编程与自动化协作","topic_url":"https://opc.beizhux.com/topics/agent-workflow","topic_path":"/topics/agent-workflow","url":"https://opc.beizhux.com/glossary/workflow","path":"/glossary/workflow","json_url":"https://opc.beizhux.com/glossary/workflow.json"},{"slug":"claude-code","name":"Claude Code","definition":"在AI觉醒星球里，「Claude Code」属于「Agent工作流」方向。聚合Agent、AI编程、Claude Code、Cursor、工作流自动化等内容，帮助你把智能体能力落到真实任务链路。 把智能体当成任务系统，而不是聊天窗口。","topic_slug":"agent-workflow","topic_name":"Agent工作流","topic_title":"Agent工作流：智能体、AI编程与自动化协作","topic_url":"https://opc.beizhux.com/topics/agent-workflow","topic_path":"/topics/agent-workflow","url":"https://opc.beizhux.com/glossary/claude-code","path":"/glossary/claude-code","json_url":"https://opc.beizhux.com/glossary/claude-code.json"}],"preview_hidden_sections":[],"related_items":[{"id":3236,"title":"FieldAI 拟融资 7 亿美元，估值达 100 亿美元","url":"https://opc.beizhux.com/content/3236/fieldai-7-100","summary":"FieldAI 正洽谈 7 亿美元融资，投后估值 100 亿美元，较上一轮 20 亿美元一年多内涨至五倍；已签署投资条款清单，领投方未披露，自 6 月以来新增至少 3,500 万美元营收及订单，总规模突破 1.35 亿美元，客户超 30 家。","access_level":"member"},{"id":3238,"title":"TypeSafe AI 决策模型 Jev 日处理量达 1 万亿 Token，OpenAI 等巨头跟进推出同类工具","url":"https://opc.beizhux.com/content/3238/typesafe-ai-jev-1-token-openai","summary":"TypeSafe AI 的决策模型 Jev 不生成文本，而是通过强化学习把输入归类到预设输出，日均处理 token 达一万亿，约 25% 的财富世界 500 强企业在用，OpenAI 等巨头开始跟进同类工具。","access_level":"member"},{"id":3240,"title":"BestBlogs早报：ChatGPT常驻智能体与Temporal故障恢复","url":"https://opc.beizhux.com/content/3240/bestblogs-chatgpt-temporal","summary":"BestBlogs 早报精讲三篇：ChatGPT 负责人 Tibo Sottiaux 谈常驻智能体、插件生态，以及减少 ChatGPT、工作模式、Codex 之间的复杂性；Temporal 演示把人工审批当作异步 API，等待审批时关闭工作进程、离线提交批准后重启即可恢复业务；Crusoe CEO 从供电交付与旧芯片…","access_level":"member"}],"usage_policy":{"summarizable":true,"preferred_url":"https://opc.beizhux.com/content/3234/google-researchers-find-a-way-to-keep-self-improving-ai-agents-from-memorizing-their-tests","private_fields_excluded":true}}