OpenAI 发布 LifeVal 基准：169 个日常任务测评智能体
OpenAI 拿 169 个真实生活任务测各家模型，发现模型老忘掉你几周前说过的话，做记忆助手的建议跑一下这个基准。
OpenAI 调研了 3,837 名美国消费者，从 148 项生活任务中筛出人们愿意付费的类别，这些任务年价值超 1.7 万亿美元。LifeVal 将其中最有价值的部分转化为 169 个智能体任务，涵盖理财、搬家、找工作、找托儿等场景，每个任务都配有评分标准、工作环境和离线网页副本，并经过三轮人工审核。成绩从 GPT-5.2 到 GPT-6 Astra 逐步上升，最常见的失败是给出了错误答案或建议，原因往往是漏掉了用户几周前聊天中提到的事实。仅在依赖过往对话的评分项上，模型胜率为 33%，而依赖已保存记忆的项目为 48%。
Recommended benchmark from OpenAI on agents for everyday life tasks.
They surveyed 3,837 US consumers to rank 148 life tasks by what people would pay an assistant to handle. Together, those tasks are worth over $1.7 trillion a year.
LifeVal turns the most valuable of them into 169 agent tasks. Each one is written by someone who has done the task many times in real life, using their own ChatGPT conversations, memories, and files. The tasks include managing finances, moving apartments, searching for a job, and finding childcare.
Every task comes with a rubric, a working environment and an offline copy of the web, and passes three rounds of human review.
Scores rise steadily from GPT-5.2 to GPT-6 Astra. The most common failure for every model is a wrong answer or recommendation, often because the model missed a fact sitting in the user's past chats. On criteria tied only to past conversations, the win rate was 33%, compared with 48% for saved memories.
If you build assistants with memory, test whether the agent actually finds what the user told it weeks ago.
https://t.co/lVuZkcx8Nt
Chat with Paper: https://t.co/TU5Ot5vcoV
