LifeSearch Shows Why Deep Research Agents Still Struggle in Real Life

Introduction

Deep research agents are designed to handle open-ended information needs through repeated web searches, multi-step reasoning, and evidence synthesis. Their outputs can look much more useful than a short chatbot response, but appearance is not the same as reliability. Existing evaluations have often emphasized academic research or synthetic tasks, leaving an important question underexplored: can an agent actually help with the messy, constraint-heavy problems people face in everyday life and professional work?

LifeSearch, introduced by the Tencent Hunyuan team, is designed to address that gap. Rather than treating research quality as a matter of fluent writing alone, it evaluates whether an agent finds the right information, covers the required points, and supports concrete claims with checkable evidence.

Key findings

• Realistic task design. The benchmark includes 203 Chinese and English tasks across seven concrete life and productivity domains. The tasks are derived from real-user query seeds and require agents to search, compare, reason across sources, and synthesize a final response.
• Fine-grained human rubrics. The authors report more than 4,500 hours of human annotation and over 4,000 expert-written rubrics. These rubrics assess both broad report quality and precise details such as values, units, dates, and URLs.
• A common evaluation setup. Twelve advanced deep research agents are tested through a unified harness. The paper also examines recurring performance patterns and failure modes instead of presenting a single leaderboard number alone.
• Substantial room for improvement. The best-performing frontier agent reported in the paper, GPT-5.6 Sol, scores 63.4%. That result indicates that a well-structured report may still omit required information, mis-handle details, or fail to align evidence with conclusions.
• Human judgment shows strong agreement. The grading configuration is selected by comparing candidate language-model judges with human rubric labels. A separate human ranking study reports a Spearman correlation of 0.89 and a Kendall correlation of 0.81 between aggregate benchmark scores and human judgments.

Why it matters

LifeSearch is important because real-world research rarely has one clean answer. A user may impose several constraints at once, while relevant information can be scattered across pages, expressed in different units, or updated over time. An evaluation that only asks whether a response sounds informative will miss the failures that matter most in practice.

The benchmark therefore shifts attention from presentation quality to evidence-grounded task completion. An agent may be strong at discovering sources and organizing prose, yet still fail at tracking every requirement, validating numbers, or maintaining consistency across a long research chain. The reported 63.4% ceiling is best read as a capability signal: current systems are useful assistants, but not yet dependable substitutes for careful human verification.

For developers, LifeSearch offers a way to diagnose weaknesses in search planning, source attribution, detail extraction, and final-answer checking. For users, it reinforces a practical rule: a long and confident report should not be trusted without checking its key claims. Future agents will need not only broader retrieval, but also better evidence-to-claim alignment and explicit checks for omissions and contradictions.

Source: Hugging Face Daily Papers