WebTraveler: A Multi-Agent Framework for Continuous Exploration in Website Traversal

ACL ARR 2026 January Submission1231 Authors

Keywords: LLM-based Agent, Web Agent, Web Traversal

Abstract: Large language models (LLMs) are increasingly deployed as web-facing agents that interact with live websites to answer complex questions. Many such questions require multi-step, in-site navigation and cross-page evidence aggregation, a regime referred to as web traversal. Existing LLM-based web agents, including recent explore–critic architectures, typically store interaction history as an ever-growing unstructured text trace and mix high-level planning with low-level page actions. This makes long-horizon behavior brittle, obscures the source of evidence, and often leads to redundant browsing or premature stopping. To mitigate these issues, we propose WebTraveler, a hierarchical multi-agent framework that separates high-level information-seeking planning from local web exploration and maintains a structured trajectory memory that compresses interaction steps into decision-relevant signals. This design allows the planner to reason over a compact, explicitly organized history, improving stability, credit assignment and cross-page evidence aggregation on deep trajectories. To obtain scalable supervision for traversal roles, we further propose a path-guided data synthesis method, which uses executable navigation paths as anchors to synthesize trajectories and applies rewriting and logical-consistency checks to ensure verifiable evidence chains. Experimental results on WebWalkerQA show that WebTraveler consistently achieves higher traversal success than previous agents across various LLM scales, with path-guided supervision bringing additional gains, especially on long-horizon tasks and in settings with larger token budgets. Codes and data collection websites will be opensourced.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: AI / LLM Agents,NLP Applications,Question Answering

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 1231