Deep Research Agents:
A Systematic Examination And Roadmap
深度研究代理： 系统检查和路线图

Recent advancements in large reasoning models (LRMs) have greatly enhanced the ability of language models to tackle complex and abstract tasks. These models have shown significant improvements in tasks such as arithmetic, common-sense reasoning, and symbolic problem-solving, largely due to innovations in model architectures and training techniques. One such advancement is Chain-of-Thought (CoT) prompting, introduced by Wei et al. wei2023chainofthoughtpromptingelicitsreasoning , which explicitly guides models to articulate intermediate logical steps, decomposing complex problems into simpler, sequential stages. This has led to notable improvements in both the interpretability and accuracy of LLMs on various reasoning benchmarks. Building upon CoT, subsequent research has introduced methods to further enhance LLM reasoning, particularly in handling lengthy textual contexts. Approaches such as positional interpolation and sparse attention mechanisms bai2024longalign ; wang2024beyond have been proposed to extend the effective context window. Furthermore, specialized benchmarks like LongBench bai2024longbench and LongFinanceQA lin2025facilitating have been developed to rigorously evaluate and improve the performance of these models in extended-context reasoning.
大型推理模型 （LRM） 的最新进展极大地增强了语言模型处理复杂和抽象任务的能力。这些模型在算术、常识推理和符号问题解决等任务方面显示出显著改进，这主要是由于模型架构和训练技术的创新。其中一项进步是 Wei 等人引入的思维链 （CoT） 提示 wei2023chainofthoughtpromptingelicitsreasoning ，它明确指导模型阐明中间逻辑步骤，将复杂问题分解为更简单的顺序阶段。这导致 LLM 在各种推理基准上的可解释性和准确性都有显着提高。在 CoT 的基础上，随后的研究引入了进一步增强 LLM 推理的方法，尤其是在处理冗长的文本上下文时。位置插值和稀疏注意力机制等方法 bai2024longalign ; wang2024 超越 已提议扩展有效上下文窗口。此外，还开发了 LongBench bai2024longbench 和 LongFinanceQA lin2025 便利 等专业基准测试，以严格评估和改进这些模型在扩展上下文推理中的性能。

To address reasoning tasks that require real-time or specialized external knowledge, frameworks like Toolformer schick2023toolformerlanguagemodelsteach and MultiTool-CoT inaba2023multitool have been proposed, enabling LLMs to autonomously incorporate external computational resources and APIs directly within reasoning workflows. These approaches effectively enhance performance in tasks dependent on precise numerical calculations and dynamic information retrieval. Maintaining reasoning coherence across multiple conversational turns also poses distinct challenges. Techniques such as Dialogue CoT chae2023dialogue and Structured CoT (SCoT) sultan2024structured explicitly integrate dialogue states and conversational contexts within reasoning chains, significantly improving coherence, context-awareness, and the ability to manage iterative interactions and clarify complex user queries. However, despite substantial improvements, existing reasoning frameworks still encounter critical issues, including hallucinations, static or outdated internal knowledge, and insufficient responsiveness to rapidly changing information needs. These limitations highlight the necessity of integrating external information sources, real-time retrieval mechanisms, and adaptive reasoning strategies, which are core motivations driving recent advances toward more comprehensive and robust reasoning frameworks suitable for DR Agent applications.
为了解决需要实时或专业外部知识的推理任务，已经提出了 Toolformer schick2023toolformerlanguagemodelsteach 和 MultiTool-CoT inaba2023multitool 等框架，使 LLM 能够自主地将外部计算资源和 API 直接整合到推理工作流程中。这些方法有效地提高了依赖于精确数值计算和动态信息检索的任务的性能。在多个对话回合中保持推理的连贯性也带来了明显的挑战。对话 CoT chae2023dialogue 和结构化 CoT （SCoT） sultan2024structured 等技术明确地将对话状态和对话上下文集成到推理链中，显著提高了连贯性、上下文感知以及管理迭代交互和澄清复杂用户查询的能力。然而，尽管有了实质性的改进，现有的推理框架仍然遇到关键问题，包括幻觉、静态或过时的内部知识，以及对快速变化的信息需求的响应不足。这些限制凸显了集成外部信息源、实时检索机制和自适应推理策略的必要性，这些是推动最新进展的核心动机，旨在实现适用于 DR 代理应用程序的更全面、更强大的推理框架。

Retrieval-augmented Generation (RAG), leveraging external knowledge bases (e.g., webs, APIs), has emerged as an effective strategy to mitigate hallucination problems and enhance the accuracy of web information search fan2024survey ; gao2023retrieval ; singh2025agentic . Early RAG architectures typically involved a static pipeline, where retrievers fetched relevant documents from external sources such as Wikipedia or search engines, and generators (e.g., LLMs) produced answers based solely on these retrieved passages. However, static approaches were limited in handling complex or multi-step queries, motivating recent advances toward iterative and interactive retrieval mechanisms to generate richer and more relevant responses, including FLARE zhang2024enhancing , Self-RAG asai2023self , IAG zhang2023iag , and ToC kim2023tree . In addition, studies izacard2023atlas ; lin2023ra expanded retrieval sources from structured databases (e.g., Wikipedia) to large-scale, diverse web corpora such as the Common Crawl dump preprocessed via the CCNet pipeline fu2022ccnet . Further improvements of RAG include hybrid approaches that combine internal LLM knowledge and external retrievals for better accuracy and coherence aliannejadi2024trec . Recently, Huang et al. huang2025rag proposed RAG-RL, introducing reinforcement learning and curriculum learning techniques, enabling reasoning language models (RLMs) to more effectively identify and utilize relevant contexts.
利用外部知识库（例如，网络、API）的检索增强生成 （RAG） 已成为缓解幻觉问题和提高网络信息搜索准确性的有效策略 fan2024 调查 ; GAO2023 检索 ; singh2025 代理 .早期的 RAG 架构通常涉及一个静态管道，其中检索器从外部来源（如 Wikipedia 或搜索引擎）获取相关文档，而生成器（例如 LLM）仅根据这些检索到的段落生成答案。然而，静态方法在处理复杂或多步骤查询方面受到限制，这推动了迭代和交互式检索机制的最新进展，以生成更丰富、更相关的响应，包括 FLARE zhang2024sulincing 、Self-RAG asai2023self 、IAG zhang2023iag 和 ToC kim2023tree 。此外，研究 izacard2023atlas ; lin2023ra 将检索源从结构化数据库（例如维基百科）扩展到大规模、多样化的 Web 语料库，例如通过 CCNet 管道 fu2022ccnet 预处理的 Common Crawl 转储。RAG 的进一步改进包括将内部 LLM 知识和外部检索相结合的混合方法，以提高准确性和连贯性 aliannejadi2024trec 。最近，Huang et al. huang2025rag 提出了 RAG-RL，引入了强化学习和课程学习技术，使推理语言模型 （RLM） 能够更有效地识别和利用相关上下文。

Despite these advancements in retrieval methods and reasoning-enhanced models, RAG approaches still face limitations in effectively managing complex reasoning workflows and dynamically adapting to varied task requirements. To address these challenges, recent research extends RAG into an agentic paradigm, integrating additional reasoning and decision-making layers atop conventional RAG pipelines singh2025agentic . Agentic RAG approaches leverage iterative retrieval, adaptive querying, and dynamic workflow adjustments, significantly enhancing multi-step reasoning capabilities. For example, RL-based query refinement techniques (e.g., Hsu et al. hsu2024grounding ) improve retrieval for complex queries, while graph-based retrieval (e.g., GeAR shen2024gear ) further enhances the processing of multi-hop queries. Despite these advancements, agentic RAG still faces critical challenges, including balancing computational overhead from dynamic reasoning processes singh2025agentic , aligning agent behaviors with user intentions zerhoudi2024personarag , and ensuring interpretability in adaptive workflows hsu2024grounding ; singh2025agentic . Moreover, even advanced agentic RAG approaches remain constrained by their reliance on pre-existing or periodically updated corpora, limiting their ability to handle real-time, rapidly changing, or long-tail information needs effectively. Addressing this challenge requires integrating external APIs and web browsing capabilities into RAG architectures, motivating recent DR methods aimed at further enhancing retrieval comprehensiveness and adaptability.
尽管检索方法和推理增强模型取得了这些进步，但 RAG 方法在有效管理复杂的推理工作流程和动态适应不同的任务要求方面仍然面临限制。为了应对这些挑战，最近的研究将 RAG 扩展到代理范式，在传统的 RAG 管道 singh2025agentic 上集成了额外的推理和决策层。代理 RAG 方法利用迭代检索、自适应查询和动态工作流调整，显著增强了多步骤推理能力。例如，基于 RL 的查询优化技术（例如，Hsu et al. hsu2024grounding ）改进了复杂查询的检索，而基于图形的检索（例如，GeAR shen2024gear ）进一步增强了多跳查询的处理。尽管取得了这些进步，代理 RAG 仍然面临关键挑战，包括平衡动态推理过程的计算开销 singh2025agentic ，使代理行为与用户意图保持一致 zerhoudi2024personarag ，以及确保自适应工作流程中的可解释性 hsu2024grounding ; singh2025 代理 .此外，即使是先进的代理 RAG 方法仍然受到对预先存在或定期更新的语料库的依赖的限制，从而限制了它们有效处理实时、快速变化或长尾信息需求的能力。应对这一挑战需要将外部 API 和 Web 浏览功能集成到 RAG 架构中，从而推动最近的 DR 方法进一步提高检索的全面性和适应性。

Model Context Protocol (MCP) and Agent-to-Agent (A2A) have been proposed to address interoperability challenges in LLM-based agent systems, enabling efficient tool access and effective multi-agent collaboration. MCP: Traditional Tool Use (TU) agents face significant challenges, including inconsistent APIs, high maintenance costs, and redundant development efforts, severely limiting interoperability across systems schick2023toolformerlanguagemodelsteach . To address these issues, Anthropic introduced the MCP, a unified communication layer allowing LLM-based agents to interact securely and consistently with external services and data sources via standardized interfaces. MCP mitigates data silo problems by providing dynamic service discovery and uniform access patterns. A2A: Google’s A2A protocol facilitates decentralized multi-agent collaboration through structured, task-oriented dialogues. Agents from diverse vendors and model architectures can discover peers, delegate responsibilities, and collaboratively manage complex tasks as equal participants google2025a2a . By abstracting agent discovery into Agent Cards, and task coordination into Tasks and Artifacts, A2A supports flexible, incremental, multi-modal workflows, ideally suited to sophisticated collaborative scenarios.
模型上下文协议 （MCP） 和代理到代理 （A2A） 已被提出，以解决基于 LLM 的代理系统中的互作性挑战，实现高效的工具访问和有效的多代理协作。 MCP：传统工具使用 （TU） 代理面临重大挑战，包括 API 不一致、维护成本高和冗余开发工作，严重限制了跨系统的互作性 schick2023toolformerlanguagemodelsteach .为了解决这些问题，Anthropic 引入了 MCP，这是一个统一的通信层，允许基于 LLM 的代理通过标准化接口与外部服务和数据源安全一致地交互。MCP 通过提供动态服务发现和统一访问模式来缓解数据孤岛问题。 A2A：Google 的 A2A 协议通过结构化、面向任务的对话促进去中心化的多代理协作。来自不同供应商和模型架构的代理可以发现对等方、委派责任，并以平等参与者的身份协作管理复杂的任务 google2025a2a 。通过将代理发现抽象到代理卡中，将任务协调抽象到任务和工件中，A2A 支持灵活的增量多模式工作流，非常适合复杂的协作场景。

MCP and A2A complement each other by clearly separating responsibilities: MCP serves as a standardized interface for accessing external tools, while A2A orchestrates collaborative agent interactions. Together, they establish a modular and scalable foundation for open, interoperable agent ecosystems, significantly enhancing the practical capabilities of AI systems in tackling complex real-world challenges.
MCP 和 A2A 通过明确划分职责来相辅相成：MCP 充当访问外部工具的标准化接口，而 A2A 则协调协作代理交互。它们共同为开放、可互作的代理生态系统建立了模块化和可扩展的基础，显著增强了 AI 系统在应对复杂的现实挑战方面的实用能力。

To enhance reasoning depth and accuracy for handling evolving tasks, DR agents employ search engines (SE) to update their knowledge through interaction with the external environment. In Table 1, we present a comparative overview of SEs, base models, and evaluation benchmarks employed by existing DR agents. The SEs can be broadly categorized into two types:
为了提高推理深度和准确性，以处理不断变化的任务，DR 代理使用搜索引擎 （SE） 通过与外部环境的交互来更新他们的知识。在表 1 中，我们提供了现有 DR 代理采用的 SE、基本模型和评估基准的比较概述。SE 大致可分为两种类型：

1. 1)

   API-Based SEs, which interact with structured data sources, such as search-engine APIs or scientific database APIs, enabling efficient retrieval of organized information.

   Report issue for preceding element
   
   1） 基于 API 的 SE，它与结构化数据源（如搜索引擎 API 或科学数据库 API）交互，从而能够高效检索有组织的信息。
2. 2)

   Browser-Based SEs, which simulate human-like interactions with web pages, facilitating real-time extraction of dynamic or unstructured content, improving the comprehensiveness of the external knowledge.

   Report issue for preceding element
   
   2） 基于浏览器的 SE，模拟与网页的类似人类的交互，促进动态或非结构化内容的实时提取，提高外部知识的全面性。

API-based retrieval is a fast, efficient, structured, and scalable method that allows DR agents to access external knowledge sources with relatively less time and computational cost. For instance, Gemini DR leverages multi-source interfaces, most notably the Google Search API and the arXiv API, to perform large-scale retrieval across hundreds to thousands of web pages, thereby significantly expanding its information coverage. Grok DeepSearch grokdeepresearch claims to ensure both the freshness and depth of its knowledge base by maintaining a continuous index via news-outlet feeds, the Wikipedia API, and X’s native interface, and by activating a query-driven agent on demand to generate targeted sub-queries and fetch relevant pages in real time. AgentLaboratory schmidgall2025agent uses the arXiv API to extract paper metadata and abstracts for automated literature reviews. AI Scientist lu2024aiscientistfullyautomated issues requests to the Semantic Scholar API to validate the novelty and citation relationships of model-generated research ideas, and CoSearchAgent gong2024cosearchagent integrates SerpApi to deliver Slack-based, real-time search responses. DeepRetrieval jiang2025deepretrieval , operating within a reinforcement-learning framework, optimizes queries against the PubMed and ClinicalTrials.gov APIs to maximize recall on biomedical tasks, and Search-o1 li2025search combines the Bing Search API with the Jina Reader API to dynamically extract and refine passages for downstream reasoning. Whilst these API-driven methods excel at structured, high-throughput data acquisition, they generally struggle when faced with deeply nested client-side JavaScript-rendered content, interactive components, or authentication barriers, thereby motivating the development of browser-based search mechanisms capable of comprehensively extracting and analyzing dynamic or unstructured information.
基于 API 的检索是一种快速、高效、结构化且可扩展的方法，它允许 DR 代理以相对较少的时间和计算成本访问外部知识源。例如，Gemini DR 利用多源接口（最著名的是 Google Search API 和 arXiv API）在成百上千个网页上执行大规模检索，从而显著扩大了其信息覆盖范围。Grok DeepSearch grokdeepresearch 声称，通过新闻媒体提要、维基百科 API 和 X 的原生界面维护连续索引，并按需激活查询驱动的代理来生成有针对性的子查询并实时获取相关页面，从而确保其知识库的新鲜度和深度。代理实验室 schmidgall2025agent 使用 arXiv API 提取论文元数据和摘要，用于自动文献综述。AI Scientist lu2024aiscientistfullyautomated 向 Semantic Scholar API 发出请求，以验证模型生成的研究思路的新颖性和引用关系，CoSearchAgent gong2024cosearchagent 集成 SerpApi 以提供基于 Slack 的实时搜索响应。DeepRetrieval jiang2025deepretrieval 在强化学习框架内运行，优化了对 PubMed 和 ClinicalTrials.gov API 的查询，以最大限度地提高对生物医学任务的召回率，而 Search-o1 li2025search 将 Bing 搜索 API 与 Jina Reader API 相结合，以动态提取和提炼段落以进行下游推理。 虽然这些 API 驱动的方法擅长结构化、高吞吐量的数据采集，但它们在面对深度嵌套的客户端 JavaScript 渲染内容、交互式组件或身份验证障碍时通常会遇到困难，从而推动开发基于浏览器的搜索机制，能够全面提取和分析动态或非结构化信息。

Browser-based retrieval provides DR agents with dynamic, flexible, and interactive access to multimodal and unstructured web content through simulated human-like browser interactions. For example, Manus AI’s browsing agent operates a sandboxed Chromium instance for each research session, programmatically opening new tabs, issuing search queries, clicking through result links, scrolling pages until content thresholds are met, filling out form elements when necessary, executing in-page JavaScript to reveal lazily loaded sections, and downloading files or PDFs for local analysis manus2025 . Although OpenAI DR, Grok DeepSearch, and Gemini 2.5 DR do not publicly disclose the implementation details of their browsing capabilities, their ability to handle interactive widgets, dynamically rendered content, and multi-step navigation strongly suggests that they too employ comparable headless-browser frameworks behind the scenes. Among open-source studies, AutoAgent yu2024auto operates within a BrowserGym environment to scroll, interact with page components, and download files when APIs are unavailable yu2024auto ; DeepResearcher zheng2025deepresearcherscalingdeepresearch employs a dedicated Web Browsing Agent that, upon receiving a browse request, processes each segment of a webpage in turn, decides whether to continue to subsequent segments based on relevance, and incrementally aggregates pertinent information into a short-term memory buffer before returning it for reasoning. While browser-based retrieval excels at capturing real-time and deeply nested content that API calls cannot reach, it also incurs greater latency, resource consumption, and complexity in handling page variability and errors, suggesting that DR agents may benefit from hybrid architectures that combine the efficiency of API-based methods with the comprehensiveness of browser-driven exploration.
基于浏览器的检索通过模拟的类似人类的浏览器交互，为 DR 代理提供了对多模式和非结构化 Web 内容的动态、灵活和交互式访问。例如，Manus AI 的浏览代理为每个研究会话运行一个沙盒 Chromium 实例，以编程方式打开新选项卡、发出搜索查询、单击结果链接、滚动页面直到满足内容阈值、必要时填写表单元素、执行页面内 JavaScript 以显示延迟加载的部分，以及下载文件或 PDF 进行本地分析 manus2025 .尽管 OpenAI DR、Grok DeepSearch 和 Gemini 2.5 DR 没有公开披露其浏览功能的实现细节，但它们处理交互式小部件、动态呈现的内容和多步骤导航的能力强烈表明，它们也在幕后使用了类似的无头浏览器框架。在开源研究中，AutoAgent yu2024auto 在 BrowserGym 环境中运行，以便在 API 不可用时滚动、与页面组件交互和下载文件 yu2024auto ;DeepResearcher zheng2025deepresearcherscalingdeepresearch 采用专用的 Web 浏览代理，在收到浏览请求后，依次处理网页的每个段，根据相关性决定是否继续后续段，并将相关信息增量聚合到短期内存缓冲区中，然后再返回进行推理。 虽然基于浏览器的检索擅长捕获 API 调用无法访问的实时和深度嵌套内容，但它在处理页面可变性和错误时也会产生更大的延迟、资源消耗和复杂性，这表明 DR 代理可能会受益于混合架构，该架构将基于 API 的方法的效率与浏览器驱动的探索的全面性相结合。

To expand DR agents’ capacity to interact with external environments in complex research tasks, specifically by actively invoking and handling diverse tools and data sources, various DR agents have introduced three core tool modules: code interpreters, data analytics, multimodal processing, along with the Model Context Protocol.
为了扩展 DR 代理在复杂研究任务中与外部环境交互的能力，特别是通过主动调用和处理各种工具和数据源，各种 DR 代理引入了三个核心工具模块：代码解释器、数据分析、多模态处理以及模型上下文协议。

As shown in Figure 4, this section systematically analyzes the construction of DR systems, focusing on workflows categorized into static and dynamic types. We then discuss planning strategies, which enhance task allocation and execution through three distinctive user interaction types to clarify intent: planning-only (direct planning without clarifying user intent), intent-to-planning (clarifying intent before planning to align the task with user goals), and unified intent-planning (generating a plan and requesting user confirmation). The distinction between single-agent and multi-agent systems is examined in the context of dynamic workflows, emphasizing specialization in task management. Additionally, we examine memory mechanisms for managing and integrating retrieved information, which enhance the performance and adaptability of DR systems.
如图 4 所示，本节系统地分析了 DR 系统的构建，重点介绍了分为 静态 和 动态 类型的工作流。然后，我们讨论了规划策略，这些策略通过 三种不同的用户交互类型 来增强任务分配和执行，以阐明意图：仅规划（直接规划而不明确用户意图）、意向到规划（在规划之前澄清意图，使任务与用户目标保持一致）和统一意图规划（生成计划并请求用户确认）。在动态工作流的上下文中研究了 单代理 系统和 多代理 系统之间的区别，强调任务管理的专业化。此外，我们还研究了用于管理和集成检索到的信息的内存机制，从而提高了 DR 系统的性能和适应性。

DR agents depend heavily on LRMs and often utilize complex hierarchical workflows. Parameter-based learning approaches such as SFT and RL encounter significant obstacles in this context, including the need to scale model parameters, manage extensive volumes of structured experience data, and design increasingly intricate training algorithms. In contrast, non-parametric continual learning approaches offer a scalable alternative: agents refine their capabilities at runtime by optimizing external memory, workflows, and tool configurations through continuous interaction with the external environment rather than by updating internal weights. This non-parametric continual learning paradigm enables efficient online adaptation with minimal data and computational overhead, making it well-suited to DR agents with complex architectures.
DR 代理严重依赖 LRM，并且经常使用复杂的分层工作流。在这种情况下，基于参数的学习方法（如 SFT 和 RL）遇到了重大障碍，包括需要扩展模型参数、管理大量结构化体验数据以及设计日益复杂的训练算法。相比之下，非参数持续学习方法提供了一种可扩展的替代方案：代理通过与外部环境的持续交互而不是更新内部权重来优化外部内存、工作流程和工具配置，从而在运行时改进其功能。这种非参数持续学习范式能够以最少的数据和计算开销实现高效的在线适应，使其非常适合具有复杂架构的 DR 代理。

Non-parametric continual learning approaches, most notably case-based reasoning (CBR), are currently a mainstream method in LLM-driven agent systems. The CBR-based method enables agents to retrieve, adapt, and reuse structured problem-solving trajectories from an external case bank dynamically. Unlike traditional RAG-based methods, which rely on static databases, CBR facilitates online contextual adaptation and effective task-level generalisation. Such flexibility underscores its potential as a scalable and practical optimization solution for DR agents with complex architecture. DS-Agent guo2024dsagentautomateddatascience is a pioneering LLM-driven agent that introduced CBR into automated data science workflows, employing approximate online retrieval from a constructed case bank. Similarly, LAM guo2025optimizing applies CBR techniques to functional test generation, combining trajectory-level retrieval with LLM planning in a modular system design. Although DS-Agent itself does not include a learning phase, Agent K grosnit2024largelanguagemodelsorchestrating advances this paradigm with dynamic external case retrieval and reuse guided by a reward-based memory policy, which exemplifies genuine self-evolution enabling continual adaptation and optimization without updating model parameters. Focusing on DR agents, AgentRxiv schmidgall2025agentrxiv further extends this paradigm by enabling autonomous research agents to collaboratively share and access a centralized repository of prior research outputs. This framework allows LLM agent laboratories to upload and retrieve reports from a shared preprint server, simulating an online-updating arXiv-like platform which can be seen as a comprehensive case bank. Such a system empowers agents to enhance their capabilities and knowledge through contextual adaptation without modifying their model parameters.
非参数持续学习方法，尤其是基于案例的推理 （CBR），目前是 LLM 驱动的代理系统中的主流方法。基于 CBR 的方法使代理能够动态地从外部案例库中检索、调整和重用结构化的问题解决轨迹。与依赖于静态数据库的传统基于 RAG 的方法不同，CBR 有助于在线上下文适应和有效的任务级泛化。这种灵活性凸显了它作为具有复杂架构的 DR 代理的可扩展且实用的优化解决方案的潜力。DS-Agent guo2024dsagentautomateddatascience 是一个开创性的 LLM 驱动代理，它将 CBR 引入自动化数据科学工作流程，采用从构建的案例库中进行近似在线检索。同样，LAM guo2025optimizing 将 CBR 技术应用于功能测试生成，在模块化系统设计中将轨迹级检索与 LLM 规划相结合。尽管 DS-Agent 本身不包括学习阶段，但 Agent K grosnit2024largelanguagemodelsorchestrating 通过动态外部案例检索和重用来推进这种范式，该模式由基于奖励的内存策略指导，它体现了真正的自我进化，无需更新模型参数即可实现持续适应和优化。AgentRxiv schmidgall2025agentrxiv 专注于 DR 代理，通过使自主研究代理能够协作共享和访问先前研究成果的集中存储库，进一步扩展了这一范式。 该框架允许 LLM 代理实验室从共享的预印本服务器上传和检索报告，模拟一个在线更新的类似 arXiv 的平台，该平台可以被视为一个全面的案例库。这样的系统使代理能够通过上下文适应来增强他们的能力和知识，而无需修改他们的模型参数。

Compared to prompt-based methods, which encode fixed demonstrations or task heuristics into static input templates, Non-parametric methods enable dynamic retrieval and adaptation of structured trajectories, thereby facilitating continual task generalization without manual prompt engineering. Relative to RAG, which typically retrieves unstructured textual content from static corpora, CBR operates at the trajectory level and emphasizes reasoning-centered memory organization. A notable example is the Kaggle Grandmaster Agent grosnit2024largelanguagemodelsorchestrating , which demonstrates how LLMs equipped with modular reasoning components and persistent memory can achieve expert-level structured problem solving, aligning closely with the CBR paradigm. These characteristics make CBR particularly well-suited for agents requiring procedural adaptation and context-sensitive optimization across tasks. Except memory-based method, self-evolution can also arise from dynamic infrastructure adaptation. For example, Alita qiu2025alita monitors task requirements and environmental signals to provision and configure new MCP servers at runtime, seamlessly extending and refining its toolset on demand.
与将固定演示或任务启发式编码为静态输入模板的基于提示的方法相比，非参数方法能够动态检索和适应结构化轨迹，从而促进持续的任务泛化，而无需手动提示工程。相对于通常从静态语料库中检索非结构化文本内容的 RAG，CBR 在轨迹级别运行，强调以推理为中心的记忆组织。一个值得注意的例子是 Kaggle Grandmaster Agent grosnit2024largelanguagemodelsorchestrating ，它演示了配备模块化推理组件和持久内存的 LLM 如何实现专家级的结构化问题解决，与 CBR 范式紧密保持一致。这些特性使 CBR 特别适合需要跨任务进行程序适应和上下文敏感优化的代理。除了基于内存的方法外，动态基础设施适应也可以产生自我进化。例如，Alita qiu2025alita 监控任务要求和环境信号，以在运行时预置和配置新的 MCP 服务器，从而按需无缝扩展和完善其工具集。

In summary, these self-evolution paradigms in LLM-driven DR agent systems offer substantial promise for structured reasoning and dynamic retrieval and open new pathways for efficient knowledge reuse and continual learning. Although these methods have not yet achieved widespread attention, they address the high data and computational demands inherent to parameter-based approaches and therefore represent an attractive direction for future research and practical deployment.
总之，LLM 驱动的 DR 代理系统中的这些自我进化范式为结构化推理和动态检索提供了巨大的前景，并为高效的知识重用和持续学习开辟了新的途径。尽管这些方法尚未得到广泛关注，但它们解决了基于参数的方法固有的高数据和计算需求，因此代表了未来研究和实际部署的一个有吸引力的方向。

OpenAI recently introduced its DR capability openai2025deepresearch , employing a single-agent architecture centered around a reinforcement learning-based, fine-tuned o3 reasoning model. Upon receiving a research query, the system initiates a concise interactive clarification step to accurately define user intent and research objectives. It then autonomously formulates and executes a sophisticated, multi-step research strategy, encompassing multimodal information retrieval, web browsing, and computational tasks such as data analysis and visualization through browser tools. Technologically, this solution delivers three significant advancements: (1) A dynamically adaptive iterative research workflow: Capable of refining its strategy throughout task execution. (2) Enhanced context memory and robust multimodal processing capabilities: Facilitating effective integration of diverse information sources. (3) Comprehensive toolchain integration: Combining web browsing capabilities with built-in programming tools to produce structured, authoritative reports supported by precise citations.
OpenAI 最近推出了其 DR 功能 openai2025deepresearch ，采用以基于强化学习、微调的 o3 推理模型为中心的单代理架构。收到研究查询后，系统会启动一个简洁的交互式澄清步骤，以准确定义用户意图和研究目标。然后，它自主制定并执行复杂的多步骤研究策略，包括多模态信息检索、网页浏览和计算任务，例如通过浏览器工具进行数据分析和可视化。在技术上，该解决方案实现了三个重大进步：（1） 动态自适应迭代研究工作流程 ：能够在整个任务执行过程中完善其策略。（2） 增强的上下文记忆和强大的多模态处理能力 ：促进不同信息源的有效集成。（3） 全面的工具链集成 ：将 Web 浏览功能与内置编程工具相结合，生成由精确引用支持的结构化、权威报告。

Google DeepMind recently introduced Gemini DR geminideepresearch , an advanced DR agent based on its multimodal Gemini 2.0 Flash Thinking model. Gemini’s reinforcement learning-driven fine-tuning, facilitated by a single-agent architecture, has been shown to enhance planning and adaptive research capabilities, enabling the system to autonomously and expeditiously complete complex tasks. Technologically, this solution delivers four significant advancements: (1) Interactive Research Planning: Upon receiving a research query, Gemini autonomously formulates a multi-step investigation plan for interactive user review and modification. (2) Asynchronous Task Management: Adopts asynchronous task management architecture to efficiently handle multiple simultaneous tasks. (3) Large-scale context windows RAG ensembles: Enabling effective management and coherent synthesis of multimodal data (eg. text, images)for in-depth professional research analysis. (4) High speed adaptive retrieval: Implements fast, multi-round adaptive web search that significantly outperforms other agents in terms of retrieval speed and amount of information per iteration.
Google DeepMind 最近推出了 Gemini DR geminideepresearch ，这是一款基于其多模态 Gemini 2.0 Flash Thinking 模型的高级 DR 代理。Gemini 的强化学习驱动的微调，由单代理架构促进，已被证明可以增强规划和自适应研究能力，使系统能够自主、快速地完成复杂的任务。在技术上，该解决方案提供了四项重大进步：（1） 交互式研究计划 ：收到研究查询后，Gemini 会自动制定一个多步骤调查计划，用于交互式用户审查和修改。（2） 异步任务管理 ：采用异步任务管理架构，高效处理多个同时进行的任务。（3） 大规模上下文窗口 RAG 集成 ：实现多模态数据（例如文本、图像）的有效管理和连贯合成，以进行深入的专业研究分析。（4） 高速自适应检索 ：实现快速、多轮自适应 Web 搜索，在检索速度和每次迭代的信息量方面明显优于其他代理。

Perplexity’s recently developed DR agent perplexitydeepresearch has demonstrated an advanced capability to decompose complex queries into well-defined subtasks. The system is capable of conducting targeted web searches iteratively, critically evaluating authoritative sources, and synthesizing structured, comprehensive reports. Technologically, this solution delivers two significant advancements: (1) Iterative Information Retrieval: Conducts successive rounds of targeted web searches with dynamic adjustments based on interim insights, ensuring comprehensive information coverage and accuracy. (2) Dynamic Prompt-Guided Model Selection: Use hybrid architecture to autonomously select the optimal combination of specialized models based on the requirements and context of specific tasks, thereby enhancing adaptability and effectiveness in various research scenarios.
Perplexity 最近开发的 DR 代理 perplexitydeepresearch 展示了将复杂查询分解为定义明确的子任务的高级功能。该系统能够迭代地进行有针对性的 Web 搜索，批判性地评估权威来源，并综合结构化、全面的报告。在技术上，该解决方案提供了两项重大进步：（1） 迭代信息检索 ：进行连续轮定向 Web 搜索，并根据临时见解进行动态调整，确保全面的信息覆盖率和准确性。（2） 动态提示引导模型选择 ：使用混合架构，根据特定任务的需求和上下文自主选择专业模型的最优组合，从而增强在各种研究场景中的适应性和有效性。

Grok DeepSearch grokdeepresearch , developed by xAI, is a computational framework that combines real-time information retrieval with multimodal reasoning to dynamically solve complex and information-rich problems. Technologically, this solution delivers two significant advancements: (1) Segment-level module processing pipline: Upon receiving a query, Grok3 initiates the credibility assessment module to identify and filter out low-quality information. Subsequently, the system’s real-time data acquisition engine gathers multimodal inputs (e.g. text, images, and code) from various sources. Subsequently, employing the sparse attention mechanism, the system undertakes key reasoning subtasks, including data cleaning, cross-source verification, and multimodal integration, in a concurrent manner. Finally, the iterative optimization process culminates in the generation of structured outputs, encompassing analysis summaries, advanced visualizations (e.g. 3D trajectories), and verifiable citations. (2) Dynamic resource allocation: Capacity for adaptively alternating between lightweight retrieval and intensive analysis modes is noteworthy, and it is further augmented by the incorporation of a secure sandbox environment for computational verification.
Grok DeepSearch grokdeepresearch 由 xAI 开发，是一个计算框架，它将实时信息检索与多模态推理相结合，以动态解决复杂且信息丰富的问题。从技术上讲，该解决方案提供了两个重大进步：（1） 分段级模块处理流水线 ：收到查询后，Grok3 会启动可信度评估模块来识别和过滤掉低质量的信息。随后，系统的实时数据采集引擎从各种来源收集多模态输入（例如文本、图像和代码）。随后，系统采用稀疏注意力机制，以并发方式承担关键推理子任务，包括数据清洗、跨源验证和多模态集成。最后，迭代优化过程以生成结构化输出而告终，包括分析摘要、高级可视化（e.g. 3D 轨迹）和可验证的引文。（2） 动态资源分配 ：在轻量级检索和密集分析模式之间自适应交替的能力值得注意，并且通过结合用于计算验证的安全沙箱环境进一步增强了它。

Microsoft recently introduced two innovative reasoning agents within Microsoft 365 Copilot: Researcher and Analyst spataro_introducing_2025 . These agents securely and compliantly access users’ work data (such as emails, meeting notes, documents, and chats) as well as web information, delivering on-demand expert knowledge.
Microsoft 最近在 Microsoft 365 Copilot 中引入了两个创新的推理代理：研究员和分析师 spataro_introducing_2025 。这些代理安全合规地访问用户的工作数据（例如电子邮件、会议记录、文档和聊天记录）以及 Web 信息，从而按需提供专业知识。

Researcher is designed to assist users in tackling complex, multi-step research tasks, delivering insights with unprecedented quality and accuracy. It combines OpenAI’s advanced research models with Microsoft 365 Copilot’s sophisticated orchestration and deep search capabilities. Users can employ Researcher to craft detailed market entry strategies, identify market opportunities for new products by integrating internal and external data, or prepare comprehensive quarterly reports for client reviews. Additionally, Researcher enhances its insights through connectors to third-party data sources such as Salesforce, ServiceNow, and Confluence.
Researcher 旨在帮助用户处理复杂的多步骤研究任务，以前所未有的质量和准确性提供见解。它将 OpenAI 的高级研究模型与 Microsoft 365 Copilot 的复杂编排和深度搜索功能相结合。用户可以聘请 Researcher 制定详细的市场进入策略，通过整合内部和外部数据来确定新产品的市场机会，或准备全面的季度报告供客户审查。此外，Researcher 还通过与第三方数据源（如 Salesforce、ServiceNow 和 Confluence）的连接器来增强其洞察。

Analyst is built as an advanced data analytics agent that rapidly transforms raw data into valuable insights within minutes. It leverages OpenAI’s o3-mini inference model, specifically optimized for advanced analytical tasks in professional environments. Analyst uses a chain-of-thought reasoning approach, solving problems step-by-step, generating high-quality responses that closely mirror human analytical thinking.
Analyst 是作为高级数据分析代理构建的，可在几分钟内将原始数据快速转换为有价值的见解。它利用 OpenAI 的 o3-mini 推理模型，专门针对专业环境中的高级分析任务进行了优化。Analyst 使用思维链推理方法，逐步解决问题，生成与人类分析思维密切相关的高质量响应。

Alibaba Qwen recently launched Qwen Deep Research, an advanced research agent powered by its flagship multimodal model Qwen3-235B-A22B. Through reinforcement learning-optimized task scheduling within a unified agent framework, the system demonstrates enhanced autonomous planning and adaptive execution capabilities, enabling rapid completion of complex research workflows. Key technological advancements include: (1) Dynamic Research Blueprinting with interactive plan refinement. (2) Concurrent Task Orchestration enabling parallel retrieval validation synthesis.
阿里巴巴 Qwen 最近推出了 Qwen Deep Research，这是一款由其旗舰多模式模型 Qwen3-235B-A22B 提供支持的高级研究代理。通过在统一的代理框架内进行强化学习优化的任务调度，该系统展示了增强的自主规划和自适应执行能力，从而能够快速完成复杂的研究工作流程。主要技术进步包括：（1） 具有交互式计划改进的 动态研究蓝图 。（2） Concurrent Task Orchestration 支持并行检索验证合成。

In addition to the pioneering DR services previously discussed, major technology corporations such as Microsoft and ByteDance, alongside emerging startups including Jina AI jina_deepsearch , H2O h2oai , and Zhipu AI zhipu2025autoglm , have also introduced their proprietary DR platforms. The advent of these solutions has spurred considerable global interest, reflected by their rapid proliferation, thereby underscoring both the technological attractiveness and substantial market potential of DR applications. Looking forward, continuous advancements in LLM reasoning, retrieval integration techniques, and multimodal generation are expected to enable DR agents to transcend traditional information retrieval and basic tool invocation tasks. Consequently, DR systems are anticipated to tackle increasingly sophisticated reasoning and complex knowledge-construction challenges, ultimately positioning DR as a foundational technological pillar for next-generation intelligent collaborative research platforms.
除了前面讨论的开创性 DR 服务外，Microsoft 和字节跳动等主要科技公司，以及 Jina AI jina_deepsearch 、H2O h2oai 和 Zhipu AI zhipu2025autoglm 等新兴初创公司也推出了他们专有的 DR 平台。这些解决方案的出现激发了全球的极大兴趣，其迅速扩散反映了这一点，从而凸显了 DR 应用的技术吸引力和巨大的市场潜力。展望未来，LLM 推理、检索集成技术和多模态生成的不断进步有望使 DR 代理超越传统的信息检索和基本工具调用任务。因此，预计 DR 系统将应对日益复杂的推理和复杂的知识构建挑战，最终将 DR 定位为下一代智能协作研究平台的基础技术支柱。