跳转至

文章背景与核心概要

大语言模型(LLM)智能体正逐渐超越传统的检索增强生成(RAG),走向允许访问整个语料库的直接语料库交互(Direct Corpus Interaction, DCI)模式。然而,由于交互预算有限,智能体往往会遭遇“证据盲区”(Evidence Blindness)——即所需的证据在静默中被漏掉、未被打开或未被阅读,导致渐进式的性能退化。

为了克服这一挑战,本文作者推出了 AtlasNav,一个持久化的多视图语料库导航框架。AtlasNav 改变了每次查询都要在线重建共享结构的传统做法,而是将语料库预先组织成一个语料库地图(Corpus Atlas),从而实现高效的自适应导航。该方法在 BrowseComp-Plus 上实现了 92.05% 的严格准确率,在降低计算成本的同时展现出了极强的可扩展性与鲁棒性。


直接语料库交互中的证据盲区:使用 AtlasNav 进行持久导航 (Evidence Blindness in Direct Corpus Interaction: Persistent Navigation with AtlasNav)

Authors: Hongyu Guo, Zhiyu Zheng, Zhao Cao
Subject: Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.24764
Submitted: August 25, 2026

Authors: Hongyu Guo, Zhiyu Zheng, Zhao Cao
Subject: Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.24764
Submitted: August 25, 2026


📌 执行摘要 (Executive Summary)

大语言模型(LLM)智能体正日益超越传统的检索增强生成(RAG),迈向直接语料库交互(Direct Corpus Interaction, DCI)阶段,在此阶段整个语料库均可被访问。然而,有限的交互预算常常导致智能体陷入证据盲区(Evidence Blindness)——这是一种渐进式的静默丢失现象,表现为所需的证据未被发现、未被打开或未被阅读。

为了克服这一问题,作者推出了 AtlasNav,一个持久化的多视图语料库导航框架。AtlasNav 没有针对每个查询在在线状态下重建共享结构,而是将语料库一次性组织为语料库地图(Corpus Atlas),从而实现自适应导航。

Large Language Model (LLM) agents are increasingly moving beyond traditional Retrieval-Augmented Generation (RAG) toward Direct Corpus Interaction (DCI), where the entire corpus remains accessible. However, limited interaction budgets often cause agents to experience Evidence Blindness—a progressive silent loss where required evidence remains undiscovered, unopened, or unread.

To overcome this, the authors introduce AtlasNav, a persistent multi-view corpus-navigation framework. Instead of rebuilding shared structures online for every query, AtlasNav organizes the corpus once into a Corpus Atlas, enabling adaptive navigation.

核心亮点:

  • 高准确率:BrowseComp-Plus 上实现了 92.05% 的严格准确率
  • 成本效益: 与最先进的动态工作空间方法相比,在线推理成本降低了 30.21%
  • 可扩展性与鲁棒性: 成功扩展至 PhantomWiki 中 10K 到 1M 规模的语料库,并能无缝迁移至异构企业知识库。

Key Highlights:

  • High Accuracy: Achieves 92.05% strict accuracy on BrowseComp-Plus.
  • Cost Efficiency: Reduces recorded online inference costs by 30.21% compared to state-of-the-art dynamic-workspace methods.
  • Scalability & Robustness: Successfully scales across 10K–1M corpora in PhantomWiki and transfers seamlessly to heterogeneous enterprise knowledge.

🔍 引言与问题陈述:证据盲区 (Introduction & Problem Statement: Evidence Blindness)

虽然直接语料库交互(DCI)保持了完整语料库的可访问性,但在有限的交互预算下,可达到的证据往往变得无法使用。这种失效在多个阶段表现为证据盲区的渐进式静默丢失: 1. 检索失败: 所需的证据根本无法浮出水面。 2. 遗漏: 浮出水面的支撑文档保持未打开状态。 3. 片段盲区: 打开的文档未能暴露出其决定性的片段。

While Direct Corpus Interaction (DCI) keeps full corpora accessible, reachable evidence frequently becomes unusable under finite interaction budgets. This failure manifests as Evidence Blindness through progressive silent loss at multiple stages: 1. Retrieval Failure: Required evidence fails to surface altogether. 2. Omission: A surfaced supporting document remains unopened. 3. Fragment Blindness: An opened document fails to expose its decisive fragment.

当前方法的局限性

  • 原始交互(Raw Interaction): 几乎没有增加可重用的语料库组织结构。
  • 动态工作空间方法(Dynamic-Workspace Methods): 从每个查询和轨迹中动态重建一个受查询条件约束的交互空间。
  • 缺点: 这两种方法都严重依赖于在在线状态下反复恢复有用的结构。

Limitations of Current Approaches

  • Raw Interaction: Adds little reusable corpus organization.
  • Dynamic-Workspace Methods: Reconstruct a query-conditioned interaction space dynamically from each query and trajectory.
  • Drawback: Both methods rely heavily on recovering useful structures online repeatedly.

🚀 AtlasNav 框架 (The AtlasNav Framework)

AtlasNav 将大规模智能体搜索重新构想为对可重用语料库结构进行有限预算的导航

  • 语料库地图(Corpus Atlas): AtlasNav 没有针对每个查询重建共享结构,而是将语料库一次性组织为一个持久的多视图语料库地图。
  • 自适应导航(Adaptive Navigation): 查询可以自适应地穿梭于预先组织好的地图中,从而大幅减少冗余的在线计算。

AtlasNav re-formulates large-scale agentic search as finite-budget navigation over reusable corpus structures.

  • Corpus Atlas: Rather than rebuilding shared structure per query, AtlasNav organizes the corpus once into a persistent multi-view Corpus Atlas.
  • Adaptive Navigation: Queries navigate the pre-organized atlas adaptively, drastically reducing redundant online computations.

📊 实验结果 (Experimental Results)

AtlasNav 在严格的基准测试中进行了评估,并展现出了显着的性能提升:

  • BrowseComp-Plus 基准测试:
  • 达到了 92.05% 的严格准确率
  • 相较于先前最先进的动态工作空间方法,在线推理成本降低了 30.21%
  • 在匹配的预算下更早地获取了所需的完整证据,并迅速收敛到模型由证据支持的经验参考。
  • 泛化与扩展:
  • PhantomWiki 独特的语料库组织下保持了有效性。
  • 成功在 10K 到 1M 项的受控范围内实现扩展。
  • 具备向异构企业知识库进行具有竞争力的迁移能力。

AtlasNav was evaluated across rigorous benchmarks and demonstrated significant performance gains:

  • BrowseComp-Plus Benchmark:
  • Reached 92.05% strict accuracy.
  • Cut online inference costs by 30.21% relative to previous dynamic-workspace state-of-the-art methods.
  • Realized complete required evidence earlier under matched budgets and rapidly converged to the model's evidence-supplied empirical reference.
  • Generalization & Scaling:
  • Maintained effectiveness under PhantomWiki's distinct corpus organization.
  • Successfully scaled across controlled ranges from 10K to 1M items.
  • Transferred competitively to heterogeneous enterprise knowledge bases.