Deep Research Survey: A Panoramic Map of Autonomous Research Agents

> Paper: Jiarun Liu et al., "Deep Research: A Survey of Autonomous Research Agents", arXiv:2508.12752v1, 2025-08

The Core Question

How many teams are working on Deep Research? How far has the field progressed? Where should it go next?

This survey serves as a panoramic map — organizing the major works, technical approaches, evaluation methods, and open challenges across the entire Deep Research field.

What the Survey Covers

The survey, from a Shandong University team, doesn't propose a new method. Instead, it systematically organizes existing work along several key dimensions:

1. Information Acquisition

• API retrieval (search engine APIs)
• Browser simulation (Playwright/Selenium)
• Hybrid approaches
• Local document/database retrieval

2. Reasoning Strategies

• Serial (step by step)
• Parallel (multiple directions at once)
• Tree/graph-shaped (multi-branch exploration)

3. Memory Management

• Short-term (in-context)
• Long-term (external storage)
• Hierarchical memory

4. Training Methods

• SFT (supervised fine-tuning)
• RL (reinforcement learning)
• Hybrid training

5. Evaluation

• Closed-domain QA (ground-truth answers)
• Open-domain research (no ground truth, human evaluation)
• Browser tasks (WebArena, WebShop)

The Feynman Perspective: Why Surveys Matter

Imagine walking into a vast library with tens of thousands of books but no idea how they relate. A survey is the library's catalog — it doesn't tell you a book's contents, but it tells you which category it belongs to, which works are related, and what the important titles are.

The survey's greatest value lies in its analytical framework. When you encounter a new Deep Research system, you can place it on the map by asking:

• How does it acquire information? (API? browser? hybrid?)
• What's its reasoning strategy? (serial? parallel? tree-shaped?)
• Does it manage memory? (short-term? long-term? hierarchical?)
• How was it trained? (SFT? RL? hybrid?)

This framework lets you compare systems rather than examine each in isolation.

Key Findings

The survey identifies several notable trends:

1. The shift from SFT to RL Early Deep Research systems relied mainly on SFT — collecting human research trajectories for the model to imitate. But new 2024–2025 systems (R1-Searcher, DeepResearcher) almost all use RL. A field consensus has formed: imitation isn't enough; models must learn to explore autonomously.

2. Evaluation is the biggest bottleneck Evaluating Deep Research is far harder than traditional NLP tasks:

• Open-domain research has no ground-truth answers
• Human evaluation is costly and subjective
• Automatic metrics (ROUGE, BLEU) can't measure research quality
• A "standard for research quality" is needed but doesn't yet exist

3. Memory management is systematically neglected Most systems handle memory crudely — stuffing retrieved content into the context window. As research chains grow to dozens or hundreds of steps, context pressure becomes an increasingly serious, overlooked bottleneck.

The Real Open Questions

The survey's most honest section outlines open problems and future directions:

• How to evaluate "research quality"? Current evaluation focuses on "is the answer correct," but real research quality also includes whether all relevant factors were identified, whether counter-evidence was considered, and whether reasoning is coherent. No good methods exist for these dimensions yet.
• How to scale to long-horizon research? Current systems run a few to a dozen steps. Human research can last hours or days. Can LLM context windows and memory management support such long-horizon work?
• How to fuse multimodal information? Research involves tables, charts, code, and video — not just text. Current systems are mostly text-only; multimodal fusion remains open.

A Critical Perspective

One question deserves deeper discussion: is the "Deep Research" label itself misleading?

Most surveyed systems do "advanced information retrieval + synthesis." They can:

• Search multiple sources
• Compare viewpoints
• Generate cited reports

But they rarely can:

• Propose original hypotheses
• Design experiments to verify them
• Derive new theories from results
• Identify fundamental questions in a field

If "looking things up and writing a report" counts as "research," are we cheapening the word? It's like calling memorizing all the bird names "understanding birds" — exactly the cargo cult Feynman's father warned about. To the authors' credit, they don't claim these systems have achieved genuine research — but the field's naming deserves more caution.

Conclusion

This survey is required reading for the Deep Research field. It doesn't give you answers — it helps you build a framework of questions.

For researchers entering the field, reading this survey is more efficient than reading 20 original papers: it shows the full landscape and each subfield's core challenges.

For industry engineers, it offers a selection framework for comparing Deep Research solutions.

> The map is not the territory. This survey is a good map — but you have to walk into the landscape yourself to see it.

---

Source: arXiv:2508.12752