Skip to content
Search

ChatGPT Deep Research

lastmod 2025-03-12created 2025-03-075 edits

Deep Research is an AI agent integrated into ChatGPT that automatically browses the web for 5 to 30 minutes and generates a cited report on a topic the user specifies.

Categories

Q&A result memo

  • Dongkuk Dangjin Rolling Mill#Deep research Q&A results

Explanation of the term Deep Research

AI labs have recently been announcing various features under the term ‘Deep Research’

  • Google released Gemini 1.5 Deep Research in December 2024, OpenAI released Deep Research in February 2025, and Perplexity released its own Deep Research shortly after
  • In addition, DeepSeek, Alibaba's Qwen, Elon Musk's xAI and others introduced Search and Deep Search features in their chatbot assistants
  • Dozens of open-source ‘deep research’ implementations have appeared on GitHub
  • This suggests that, like Retrieval-Augmented Generation (RAG) in 2025, the term ‘deep research’ is being used without a clear definition
  • Google : “Deep Research uses AI to explore complex topics and deliver comprehensive, easy-to-read reports, showing that Gemini is getting better at handling complex tasks to save you time.” -
  • OpenAI : “Deep research is OpenAI's next agent: you give it a prompt, and ChatGPT finds, analyzes, and synthesizes hundreds of online sources to create a comprehensive report at the level of a research analyst.”
  • Perplexity : “When you ask a Deep Research question, Perplexity performs dozens of searches, reads hundreds of sources, and reasons through the material to autonomously deliver a comprehensive report.”

Setting the marketing terms aside, deep research can be defined as follows: 

사용자 쿼리를 받아들이고, 대형 언어 모델(LLM)을 에이전트로 사용하여 반복적으로 정보를 검색하고 분석하며, 상세한 보고서를 출력하는 보고서 생성 시스템

In natural language processing (NLP) terms, it is known as ’report generation’

Implementation approaches

  • Since the arrival of ChatGPT, report generation, or ‘deep research’, has become a major focus of AI engineering
  • The author experimented with it at a hackathon in early 2023, when AI engineering was just emerging
  • Tools such as LangChain, AutoGPT, GPT-Researcher, and prompt engineering, along with countless demos, drew a lot of attention on Twitter and LinkedIn
  • However, the real challenges lie in the implementation details
  • Below, we explore common patterns for building report generation systems, highlight their differences, and classify the offerings of various vendors

Untrained: directed acyclic graph (DAG)

  • Early on, it was found that asking an LLM such as GPT-3.5 to generate a report from scratch was not practical
  • Instead, the Composite pattern was used to chain multiple LLM calls
  • Decompose the user query to generate a report outline
  • For each section, retrieve relevant information from a search engine or knowledge base and summarize it
  • Finally, use an LLM to combine the sections into a coherent report
  • GPT-Researcher is an example
    • Every prompt in this system is carefully tuned through ‘prompt engineering’
    • Evaluation relies on subjectively checking the output, and report quality is inconsistent
    • It is great when it works, but it is not always reliable

Untrained: finite state machine (FSM)

  • To improve report quality, engineers added complexity to the DAG approach
  • Instead of a single-pass process, they introduced structural patterns such as Reflexion and self-reflection so that the LLM reviews and improves its own output
  • This turns the simple DAG into a finite state machine (FSM), with the LLM partly guiding the state transitions
    • As with the DAG approach, all prompts are handwritten and evaluation is subjective
    • Because the system is hand-tuned, report quality still varies widely

Trained: end-to-end

  • The drawbacks of the previous methods, haphazard prompt engineering and the lack of measurable evaluation metrics, drove the search for change
  • Stanford's STORM solved these problems by optimizing the system end to end with DSPy
    • As a result, STORM generates reports with quality comparable to Wikipedia articles

Trained: large reasoning models

  • Improvements in LLM reasoning ability have made large reasoning models an attractive option for deep research
  • For example, OpenAI trained its deep research model as follows
    • Evaluating outputs using LLM-as-a-judge and evaluation rubrics
  • Google's Gemini and Perplexity's chat assistant also offer ‘deep research’ features, but they have not published documentation on how they optimized their models or systems, or any substantive quantitative evaluation
  • However, Google's deep research product manager said in a podcast interview, “We have special access. It's pretty much the same model (Gemini 1.5). Of course, we do our own post-training work”
  • This suggests that the fine-tuning work is not substantial
  • Meanwhile, xAI's Grok excels at report generation but does not appear to search beyond two iterations
  • It searches the outline section a few times and each section a few times

Competitive landscape

  • A conceptual map was developed to evaluate the capabilities of various services offering deep research features
  • Vertical axis: depth of research (number of iterative cycles that gather additional information based on previous results)
  • Horizontal axis: level of training (from hand-tuned systems to fully trained systems using machine learning techniques)
  • Representative trained systems:
    • OpenAI Deep Research: a reinforcement learning-based system optimized for research tasks
    • DeepSeek: trained for general reasoning and tool use, and adaptable to research needs
    • Google Gemini: a broadly trained LLM, not specialized for research
    • Stanford STORM: a system that optimizes the entire research process end to end
  • This framework helps understand how each service balances the depth of iterative research against its training approach

Conclusion

  • Deep research technology is advancing rapidly, and techniques that were ineffective or unimplemented a few months ago are now being applied successfully
  • However, the ambiguous use of terms adds to the confusion
  • Hopefully this post clarifies the technical differences and helps you avoid being swayed by marketing terms

See also

Favorite site

author Zer0 sitemap