LifeSearch: best deep research agent scores only 63.4% on life tasks

Researchers at Tencent's LLM Department released LifeSearch, a benchmark of 203 Chinese and English deep-research tasks drawn from real-user seeds across seven life and productivity domains, with more than 4,000 expert rubrics built over more than 4,500 hours of annotation. Under a shared harness, the best of 12 deep-research agents, GPT-5.6 Sol, scored only 63.4%. Claude Opus 5 was next at 56.1%, and the rest were below 50%. Common failures included stopping early to ask for clarification, using evidence outside the requested time window, and leaving out specific things the task asked for. Human rankings lined up with the scores (Spearman's rho 0.89). It is a preprint posted to arXiv on 13 Oct 2026 and not yet peer reviewed. Sorted take: polished research reports still miss about a third of what was asked, so give agents explicit date ranges and required fields, and check those before you use the output.