DeepResearch: A Survey of LLM-based Research Agents

Abstract

—Large Language Models (LLMs) have demonstrated remarkable advances in natural language processing and content generation, yet they remain limited in supporting complex research tasks due to hallucinations, shallow summarization, weak multi-step reasoning, and unverifiable outputs. To address these challenges, the concept of DeepResearch has recently emerged, referring to research-oriented agents built on long-context LLMs and augmented with capabilities such as multi-step reasoning, deep retrieval, grounding and citation, agentic orchestration, and structured generation. These mechanisms enable models to act as autonomous researchers, capable of planning, retrieving, and producing systematic, evidence-based reports. Representative systems include Google Gemini Deep Research and OpenAI Deep Research, while in China platforms such as Kimi, Baidu Wenxin, Doubao, and DeepSeek are actively developing localized solutions. This paper provides a taxonomy of core techniques, presents a layered system architecture and architectural paradigms, reviews representative implementations and applications, and highlights open challenges and future directions. By consolidating current progress, we aim to guide the development of reliable and trustworthy DeepResearch agents.