Skip to content
Barry Han
Go back

WebNavigator: Global Web Navigation via Interaction Graph Retrieval

Xuanwang Zhang , Yuteng Han , Jinnan Qi , Mulong Xie , Zhen Wu , Xinyu Dai

arXiv Preprint, March 20, 2026

Autonomous agent Web navigation Multimodal retrieval Multi-site task
PDF arXiv Code Homepage

Abstract

Despite significant advances in autonomous web navigation, current methods remain far from human-level performance in complex web environments. We argue that this limitation stems from Topological Blindness, where agents are forced to explore via trial-and-error without access to the global topological structure of the environment. To overcome this limitation, we introduce WebNavigator, which reframes web navigation from probabilistic exploration into deterministic retrieval and pathfinding. WebNavigator constructs Interaction Graphs via zero-token cost heuristic exploration offline and implements a Retrieve-Reason-Teleport workflow for global navigation online. WebNavigator achieves state-of-the-art performance on WebArena and Online-Mind2Web. On WebArena multi-site tasks, WebNavigator achieves a 72.9% success rate, more than doubling the performance of enterprise-level agents. This work reveals that Topological Blindness, rather than model reasoning capabilities alone, is an underestimated bottleneck in autonomous web navigation.


The Problem: Topological Blindness

Current web agents navigate like tourists without a map — they can only see the current page and must guess where to click next. We call this Topological Blindness: agents lack access to the global structure of websites, forcing inefficient trial-and-error exploration.

Overview of WebNavigator Figure 1: WebNavigator’s two-phase paradigm. Offline: build an Interaction Graph via zero-token heuristic exploration. Online: Retrieve-Reason-Teleport workflow for global navigation.

Our Solution: WebNavigator

WebNavigator transforms web navigation from probabilistic exploration into deterministic retrieval and pathfinding.

Phase 1: Offline Interaction Graph Construction

Before any task, we build a complete map of the website:

Phase 2: Online Retrieval-Augmented Navigation

During task execution, the agent uses a Retrieve-Reason-Teleport workflow:

  1. Retrieve: Given a navigation query, find top-k relevant pages from the pre-built graph
  2. Reason: A multimodal LLM selects the best target from candidates
  3. Teleport: Compute shortest path and execute — zero token cost

Trajectory comparison Figure 2: On a multi-site task requiring cross-domain knowledge (CMS → Map), WebNavigator achieves human-level planning in 2 actions, while ReAct fails due to Topological Blindness.

Key Results

BenchmarkWebNavigatorPrevious SOTAImprovement
WebArena (overall)63.3%45.7%+38%
WebArena (multi-site)72.9%35.4%+106%
Online-Mind2Web52.7%41.0%+29%

Why It Works

  1. Complete environmental knowledge — The Interaction Graph captures the full observation space
  2. Compact action space — Only 6 actions needed (vs. 15+ in prior work)
  3. Task simplification — Transforms trajectory generation into candidate selection

Share this post on: