LLM Compass

English

简体中文

English

简体中文

Appearance

Deep Research (Deep Research Agents) Overview ​

In one sentence: Deep Research refers to a class of autonomous agents — given a research-style question, they independently complete an iterative loop of "plan → multi-step web search → read → reflect and re-search → synthesize → write up with citations," producing a verifiable long-form report; it is not a one-shot connected-to-the-web Q&A, but treats "doing research" itself as an agent trajectory that may run for several minutes to tens of minutes. The representative starting points are OpenAI Deep Research (2025-02) and Google Gemini Deep Research (2024-12), while on the open-source side there are GPT Researcher, HF open-deep-research (2025-02), Stanford STORM (2024), and others.

Prerequisite reading: Agent Overview, Tool Use, Multi-Agent; for the boundary with retrieval augmentation see Skills vs RAG/Fine-tuning; for the RL used to train these browsing agents see RL for Long-Horizon Web Navigation Agents.

Definition and Capability Boundaries ​

The term "Deep Research" was adopted almost simultaneously by several vendors from late 2024 to early 2025, with broadly consistent meaning: the user poses an open-ended question that requires synthesizing multiple sources ("compare the inference-chip roadmaps of three cloud vendors over the past three years," "investigate the latest first-line therapies for a disease and their evidence levels"), and the agent autonomously decides what to search, what to read, what is still missing, and when to wrap up, finally delivering a structured report with inline citations.

Its capability boundaries are drawn by three things:

  • The goal is "synthesis," not "looking up a single fact". A single fact ("who is a company's CEO") is fine with ordinary connected-to-the-web Q&A; the value of Deep Research lies in cases that require cross-comparing dozens to hundreds of web pages, reconciling conflicting claims, and organizing them into a long, argumentative piece.
  • It plans and iterates on its own. Rather than a fixed retrieve-then-read pass, it discovers knowledge gaps while reading, then generates new sub-questions to fill them, looping for several rounds.
  • It produces a verifiable report with citations, not an unsourced answer — citations are the core contract that distinguishes Deep Research from an ordinary "chat that can browse the web" (although citation reliability itself remains an open problem, see below).

A Typical Agent Loop ​

The vast majority of Deep Research systems (whether closed- or open-source) converge to the same iterative paradigm: plan first, then loop through "search — read — reflect" for several rounds, and finally synthesize the writeup with citations attached.

The differences among systems mainly lie in three places: whether the loop is driven by a single agent self-reflecting (GPT Researcher, HF open-deep-research) or by a supervisor dispatching multiple parallel sub-agents (LangChain open_deep_research, most products); whether browsing uses plain-text web pages or a real browser with vision; and whether the underlying model has undergone dedicated training for browsing and reasoning (OpenAI Deep Research is specifically optimized on the o-series).

Differences from RAG / Ordinary Connected-to-the-Web Q&A ​

Ordinary RAG / Web Q&ADeep Research
Number of retrievalsUsually one round (retrieve → read → answer)Multiple iterative rounds, re-searching while reading
Built-in planningGenerally noneYes, decompose sub-questions / make an outline first
Reflective re-searchNoneYes, search again when gaps are found
Output formA single answerStructured long report + inline citations
Time costSecondsMinutes to tens of minutes
Failure modePoor answer if retrieval missesAny misstep in the long trajectory accumulates bias

In short: RAG "fetches material once to answer," while Deep Research "verifies repeatedly like a researcher before writing up." Deep Research usually contains RAG internally (each round of search-read is a RAG pass), but wraps it in an agent loop of planning and reflection.

Evolution Timeline ​

Taxonomy Comparison Table ​

NameYearClosed/OpenOrgOne-linerLink
GPT Researcher2023-07OpenAssaf Elovic (community)planner+execution dual-agent, can plug in any LLM, deep mode ~5 min/reportGitHub
STORM / Co-STORM2024OpenStanford OVALWrites Wikipedia-style long-form: pre-writing research to make an outline, then writeup with citations (Co-STORM 2024-09)Details ·GitHub
Gemini Deep Research2024-12Closed/ProductGoogleReleased 2024-12, produces a research plan for user confirmation before executingBlog
OpenAI Deep Research2025-02Closed/ProductOpenAISpecifically optimizes browsing and reasoning on the o-series, released 2025-02, HLE 26.6%Details ·Blog
Perplexity Deep Research2025-02Closed/ProductPerplexityReleased 2025-02, emphasizes speed (most tasks under 3 min), HLE 21.1%Blog
Grok DeepSearch2025-02Closed/ProductxAILaunched with Grok 3 (2025-02), can integrate X real-time data and reconcile conflicting claimsRelease
open-deep-research (HF)2025-02OpenHugging FaceReproduced the OpenAI version in 24 hours, a code agent based on smolagents, GAIA validation set 55%Details ·Blog
open_deep_research (LangChain)2025OpenLangChainSupervisor architecture based on LangGraph, dispatching parallel sub-agentsGitHub
Tongyi DeepResearch2025-10OpenAlibaba Tongyi Lab30B-A3B MoE, agentic mid/post-training, open-source benchmark-topping standard (BrowseComp 43.4 / HLE 32.9 / GAIA 70.9, per the technical report)Details ·arXiv
REDSearcher2026-02OpenXiaohongshu RED · HIT · SJTUTargets sparse trajectories/rewards, low-cost unified training of long-horizon search agents, includes a multimodal version (BrowseComp 57.4 / GAIA 80.1, per the paper)Details ·arXiv
MiroFlow / MiroThinker2026-02OpenMiroMindHighly robust open-source deep-research framework + research agent model, SOTA-level on GAIA / BrowseComp(-ZH) / HLE / xbench (per the papers)arXiv ·Model
Marco DeepResearch2026-03PaperAlibaba International AIDCverification-centric: three-layer verification across data synthesis / trajectories / test-time, emphasizing efficiencyarXiv
O-Researcher2026-01OpenAcademic teamMulti-agent distillation + agentic RL, achieves competitive deep-research scores without relying on closed-source data/modelsarXiv
Mind DeepResearch2026-04PaperLi Auto~30B, three agents for planning/deep-search/report + a four-stage pipeline of SFT→Search-RL→Report-RL→preference alignment, deployed in Li Auto productsDetails ·arXiv
AgentDisCo2026-05PaperJin et al.Models deep research as adversarial optimization of exploration vs exploitation: decoupled Critic/Generator collaboration + code-generation meta-optimization accumulating a policy bank, matching closed-source on multiple report benchmarksDetails ·arXiv
DR-Rubric2026-05OpenFudan / Xiaohongshu etc.Treats "crafting RL reward rubrics" as a deep-research task: agentic retrieval to mine evidence → distill into atomic verifiable constraints → GRPO, small models can bootstrapDetails ·arXiv

The Domestic and Open-Source Benchmark Race (2025–2026) ​

From the second half of 2025 into 2026, Deep Research evolved from "a few product launches" into a head-to-head contest on public leaderboards, with the main battlegrounds being BrowseComp / BrowseComp-ZH (deep browsing to find information), HLE (hard knowledge), GAIA (general assistant), xbench-DeepSearch / WebWalkerQA / FRAMES, and others. The most representative work in this wave — those with papers and high scores — comes almost entirely from domestic large companies and the open-source community:

  • Tongyi DeepResearch (Alibaba Tongyi Lab, arXiv:2510.24701): 30B-A3B MoE, end-to-end trained with agentic mid-training + agentic post-training plus a fully automated data-synthesis pipeline, pulling open-source deep-research agents into the same tier as OpenAI Deep Research; reports BrowseComp 43.4 / BrowseComp-ZH 46.7 / HLE 32.9 / GAIA 70.9 / xbench-DeepSearch 75.0 (per the technical report). See the Tongyi DeepResearch page.
  • REDSearcher (Xiaohongshu RED × HIT × SJTU, arXiv:2602.14234): directly tackles the bottleneck of "extremely sparse high-quality search trajectories and reward signals," using a low-cost unified pipeline of complex task synthesis (graph topology + scattered evidence) + two-stage mid-training + SFT/Agentic RL, and doing rollouts on a local closed corpus of tens of millions of documents to save cost; the 30B-A3B reports BrowseComp 57.4 / GAIA 80.1 (per the paper), and extends to the multimodal REDSearcher-MM. See the REDSearcher page.
  • MiroFlow / MiroThinker (MiroMind, arXiv:2602.22808 / arXiv:2511.11793): MiroFlow is a highly robust open-source deep-research framework (agent-graph orchestration + optional deep-reasoning mode), MiroThinker is the accompanying model (scaling along model / context / interactive dimensions); achieves open-source SOTA-level scores across GAIA, BrowseComp-EN/ZH, HLE, xbench-DeepSearch and other leaderboards (exact scores vary by version, per the respective papers).
  • Mind DeepResearch (Li Auto, arXiv:2604.14518): ~30B, decomposes deep research into three agents for planning / deep-search / report, polishing search and report-writing abilities separately with a four-stage pipeline of SFT cold-start → Search-RL → Report-RL → preference alignment, and builds its own MindDR Bench benchmark of 500 Chinese queries with multi-dimensional rubrics; deployed in Li Auto's own products. Reports BrowseComp 42.8 / BrowseComp-ZH 45.7 / xbench-DS 75.0 (per the paper). See the Mind DeepResearch page.
  • Marco DeepResearch (Alibaba International AIDC, arXiv:2603.28376): threads verification through data synthesis, trajectory construction, and test-time as three layers — letting the agent act as its own verifier, suppressing the downstream propagation of errors at each stage, emphasizing "efficient deep research."
  • O-Researcher (arXiv:2601.03743): uses two-stage training of multi-agent distillation + agentic RL to let open-source models of various sizes achieve competitive deep-research scores without relying on closed-source data/models.

Leaderboard caveat: these scores come from the respective papers/technical reports, and the definitions, time points, tool configurations, and test-time scaling (e.g., Heavy Mode / parallel sub-agents) all differ, so directly comparing absolute values across systems is of limited meaning; the leaderboards (especially BrowseComp-ZH and xbench-DeepSearch) are also iterating quickly. Treat them as a "rough order-of-magnitude reference for open-source deep-research capability in the same period," not a precise ranking.

Evaluation and Current State ​

These systems mainly compete on three benchmarks:

  • GAIA (arXiv:2311.12983, 2023-11, Meta et al.): 466 real-world tasks that are "easy for humans, hard for AI," requiring reasoning, multimodality, web browsing, and tool use. HF open-deep-research reaches ~55% on the validation set (pass@1), and its blog cites OpenAI Deep Research's ~67% average as a comparison.
  • Humanity's Last Exam (HLE): hard knowledge questions across over a hundred disciplines. OpenAI Deep Research officially reports 26.6%, Perplexity Deep Research reports 21.1% — these are self-reported numbers with differing definitions and time points, so cross-comparison should be cautious.
  • BrowseComp (arXiv:2504.12516, 2025-04, OpenAI): 1266 browsing questions that require "persistent digging through deeply buried information," with short answers that are easy to auto-grade, specifically measuring an agent's ability to "dig up" information on the web.

Current state and limitations (this is also the part most worth being wary of right now):

  • Hallucination and unreliable citations: reports read authoritatively, but the claims are not necessarily supported by their attached citations — there are cases of "a link is attached but doesn't match the source text" or over-extrapolation; the presence of a citation ≠ a correct citation.
  • Uneven source quality: the agent struggles to reliably distinguish authoritative sources from marketing/low-quality content, and tends to write one-sided or outdated information into its conclusions.
  • Cost and latency: a report easily takes minutes to tens of minutes and invokes heavy retrieval and long-context reasoning; open-source solutions (e.g., GPT Researcher self-reports ~$0.4/report under o3-mini) are relatively controllable, but product-grade deep research is still expensive.
  • Poor reproducibility: web content changes over time and search results are nondeterministic, so two runs of the same question may give different results, posing difficulties for evaluation and auditing.

Practical advice: treat Deep Research as a first-draft research assistant that lists its sources, not as a final reviewer, and always trace important conclusions back to the original text via their citations.

References ​