正文 · AI 翻译
IT Home reported on November 3 that OpenAI released a new benchmark called LifeVal on November 2 to measure whether AI models can truly help ordinary people with everyday tasks. The company said the evaluation aims to bring general-purpose AI closer to providing “the everyday help that is most useful to people.”
To design LifeVal’s tasks, OpenAI first consulted the U.S. American Time Use Survey and the cross-national Multinational Time Use Study. It then surveyed 3,995 demographically representative respondents, asking which tasks they most wanted a personal AI assistant to help with.
Accordingly, LifeVal covers caregiving, health, education, household administration, personal finance, shopping, and access to government services. Typical tasks include choosing and enrolling a child in a suitable preschool after a move, making a meal and care plan for a family member with diabetes, buying groceries within a budget, organizing government forms, and coordinating community flood-prevention volunteers.
Unlike traditional benchmarks, LifeVal focuses more on whether a task can be advanced to a state where it is “actionable and ready to deliver.” For example, rather than simply listing preschools, the AI must narrow the options based on location, budget, commute, and family needs; when handling finances, it must also account for bills, income, and risks. OpenAI said models may need to understand chat histories, browse the web, combine information from multiple sources, and offer advice suited to a person’s circumstances.
OpenAI said it has used LifeVal over the past year to track and drive improvements in models’ performance on everyday tasks. Public examples show relatively high scores on some tasks: 93.1% for finding a preschool for a child, 98.0% for organizing flood-prevention volunteers, and 90.4% for preparing a meal plan for a parent with diabetes. IT Home has included the relevant screenshots below:
However, LifeVal remains primarily an internal OpenAI evaluation. The full task set, scoring code, margins of error, and an independently reproducible leaderboard have not yet been made available to outsiders, so its results should still be interpreted with caution.