LifeVal is an evaluation suite designed to measure how well large language models perform tasks people commonly face in daily life. The research publication frames LifeVal as a way to track progress on practical workflows — from finding childcare to navigating government services — and to assess models’ ability to take actions, not only offer advice.
Method and scope
The report says LifeVal builds on national and international time-use datasets, specifically the American Time Use Survey and the Multinational Time Use Study, and supplements them with a new demographically representative survey of 3,995 respondents. Tasks were selected where survey participants indicated they would most value assistance: caregiving, health, education, household administration, personal finances, shopping, and accessing government services.
Results and examples
According to the publication, model performance improved across several releases. Average unweighted task rewards rose over the past year from about 49.3% for earlier models to 80.6% for GPT-6 Astra. The timeline in the report lists model release dates and corresponding gains: Dec 11, 2025; Mar 5, 2026; Apr 23, 2026; Jul 9, 2026; and Sep 3, 2026.
The authors provide workflow examples with recorded success rates: enrolling a child in kindergarten after moving (93.1%), organizing volunteers for flood preparation (98.0%), finding a new home meeting specific constraints (100%), setting up a remote viewing session with a relative (96.3%), and creating a meal plan for a parent managing a new diabetes diagnosis (90.4%). A selection of other tasks in the report also recorded 100% success in the tested runs.
The publication notes that improvements were driven in part by better browser use, computer use, and connector integrations, which expand the range of tasks the models can complete autonomously. The authors present LifeVal as a measurement tool to guide further work on models’ ability to help with everyday decisions and responsibilities.
The LifeVal report was published on November 2, 2026.
Original source: OpenAI News