LifeSearch: Benchmarking Deep Research Agents for Real-World Life and Productivity
The benchmark says frontier deep-research agents still miss a lot on everyday work tasks.
LifeSearch tests 12 deep-research agents on 203 Chinese and English tasks drawn from real-user query seeds across seven life and productivity domains. The authors built more than 4,000 expert-written rubrics after over 4,500 hours of human annotation. The best reported agent, GPT-5.6 Sol, scored 63.4%, which the paper frames as substantial remaining headroom. The grading setup was checked against human rubric labels, and aggregate scores aligned strongly with human rankings. HF Daily Papers' note
LifeSearch tests 12 deep-research agents on 203 Chinese and English tasks drawn from real-user query seeds across seven life and productivity domains. The authors built more than 4,000 expert-written rubrics after over 4,500 hours of human annotation. The best reported agent, GPT-5.6 Sol, scored 63.4%, which the paper frames as substantial remaining headroom. The grading setup was checked against human rubric labels, and aggregate scores aligned strongly with human rankings. HF Daily Papers' note
score 5