IT之家 November 3 news: OpenAI released a new benchmark called LifeVal on November 2, intended to measure whether AI models can truly help ordinary people complete everyday tasks; the company says this evaluation aims to bring general AI closer to “the everyday help that is most useful to people.”
In designing LifeVal’s tasks, OpenAI first referred to the U.S. “American Time Use Survey” and the cross-national “Multinational Time Use Survey,” and then conducted a survey of 3,995 demographically representative respondents, asking them on which matters they most wanted a personal AI assistant to help.
Accordingly, LifeVal covers scenarios such as caregiving, health, education, household administration, personal finance, shopping, and accessing government services. Typical tasks include: after moving, selecting and enrolling a child in a suitable kindergarten; creating a diet and care plan for a family member with diabetes; purchasing ingredients within a budget; organizing government forms; and arranging community flood-prevention volunteers.
Unlike traditional benchmarks, LifeVal focuses more on whether tasks can be advanced to an “executable, deliverable” state. For example, AI should not merely list kindergartens but filter them based on location, budget, commuting, and family needs; when handling finances, it should also take into account bills, income, and risks. OpenAI says models may need to understand chat records, browse the web, integrate information from multiple sources, and give advice suited to an individual’s circumstances.
OpenAI says that over the past year it has used LifeVal to track and drive improvements in models on everyday tasks. Public examples show high performance on some tasks: finding a kindergarten for a child scored 93.1%, organizing flood-prevention volunteers scored 98.0%, and preparing a meal plan for a parent with diabetes scored 90.4%. IT之家 attaches the relevant screenshots below:
However, LifeVal is currently still mainly an internal OpenAI evaluation. The outside world has not yet obtained the complete task set, scoring code, margin of error, or an independently reproducible leaderboard, so its results still need to be interpreted with caution.
