Tencent Hunyuan Releases LifeSearch: The First Real-World Life & Productivity Benchmark for Deep Research Agents, 4,500 Hours of Human Annotation Shows Frontier SOTA at Just 63.4%

Background and the Problem

With the rapid emergence of autonomous web exploration systems such as OpenAI Deep Research and Gemini Deep Research, artificial intelligence has advanced from single-turn retrieval into long-horizon Deep Research (DR) agents capable of multi-step search orchestration, contradictory evidence verification, and extensive synthesis. However, existing agent benchmarks—such as GAIA, SWE-bench, and Humanity's Last Exam—predominantly focus on academic paper reasoning, competitive code generation, or synthetic trivia. They systematically overlook high-stakes personal decision-making and workplace productivity workflows, such as synthesizing family travel itineraries across volatile flight policies, auditing medical insurance riders, or evaluating complex enterprise SaaS procurement tiers. In these realistic domains, agents routinely hallucinate critical parameters (prices, refund dates, specific policy clauses) while masquerading behind superficially polished report structures.

Architecture and How It Works

To address this critical evaluation gap, researchers from Tencent Hunyuan developed LifeSearch, the first comprehensive benchmark engineered explicitly for real-world life and professional productivity deep research:

1. Real-User Seeded Multi-Domain Coverage: Spanning 203 demanding bilingual (English and Chinese) research tasks across seven essential domains: travel planning, personal finance, healthcare navigation, hardware/software procurement, career advancement, legal analysis, and local utility operations;
2. 4,500+ Hours of Human Curation with 4,000+ Fine-Grained Rubrics: Over 4,500 hours of domain expert annotation produced a dual-layered evaluation rubric encompassing more than 4,000 verified constraints. The harness evaluates overarching synthesis quality while enforcing rigid ground-truth verification on exact numerical values, currency units, effective dates, and authentic source URLs;
3. Rigorous Human-Aligned Grading Harness: By calibrating candidate LLM evaluator configurations directly against human grading labels, LifeSearch achieves exceptional correlation with human expert consensus (Spearman's rho = 0.89, Kendall's tau = 0.81), establishing a trustworthy testbed for frontier agent architectures.

Benchmarks and Measured Results

Systematic experiments benchmarking 12 leading deep research agents under a standardized execution environment uncovered stark performance bottlenecks:

• Frontier SOTA Capped at 63.4%: The highest-performing frontier agent, GPT-5.6 Sol, achieved an aggregate score of just 63.4%, demonstrating that current agentic architectures struggle severely when subjected to rigorous real-world constraint synthesis;
• Severe Collapse on Concrete Value Verification: While most agents achieved respectable scores (>75%) on generic report readability and broad topical coverage, scores plummeted below 40% on concrete value rubrics—revealing extensive fabrication of pricing deadlines, flight transfer feasibility, and obsolete vendor terms;
• Premature Search Cessation: Error trace analyses demonstrated that when confronted with contradictory search snippets, commercial agents routinely halted search exploration prematurely instead of methodically querying primary domain sources.

Getting Started for Developers

The LifeSearch benchmark harness and rubric schemas are maintained across Tencent Hunyuan open-source repositories (https://github.com/Tencent-Hunyuan):

• Agent Harness Integration: Developers building agentic research stacks can utilize LifeSearch rubrics to construct automated reward models for RLVR or offline trajectory filtering;
• Counter-Hallucination Guardrails: Engineering teams should implement mandatory citation verification filters that extract all quantitative assertions, dates, and prices from draft reports and verify them against cached primary HTML sources before final delivery;
• Bilingual Evaluation Support: By leveraging paired Chinese and English tasks, teams deploying consumer-facing concierge agents can systematically assess cross-lingual localization fidelity and local platform API retrieval robustness.