Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Nested Browser-Use Learning for Agentic Information Seeking

Baixuan Li, Jialong Wu, Wenbiao Yin, Kuan Li, Zhongwang Zhang, Huifeng Yin, Zhengwei Tao, Liwen Zhang, Pengjun Xie, Jingren Zhou, Yong Jiang

TL;DR

NestBrowse tackles the gap between IS agents' API-level tool use and full browser-based information gathering. It introduces a minimally complete four-action browser toolkit and a nested outer-loop/inner-loop framework, trained via multi-task imitation learning to jointly optimize reasoning and intra-page exploration. On four challenging deep IS benchmarks across English and Chinese, NestBrowse—especially the 30B-A3B model—achieves strong results and outperforms many open-source baselines while rivaling several proprietary systems, with the 4B model also competitive. The approach demonstrates that principled browser abstractions and hierarchical reasoning enable small models to perform complex deep web information seeking efficiently, reducing context waste by injecting only goal-relevant content.

Abstract

Information-seeking (IS) agents have achieved strong performance across a range of wide and deep search tasks, yet their tool use remains largely restricted to API-level snippet retrieval and URL-based page fetching, limiting access to the richer information available through real browsing. While full browser interaction could unlock deeper capabilities, its fine-grained control and verbose page content returns introduce substantial complexity for ReAct-style function-calling agents. To bridge this gap, we propose Nested Browser-Use Learning (NestBrowse), which introduces a minimal and complete browser-action framework that decouples interaction control from page exploration through a nested structure. This design simplifies agentic reasoning while enabling effective deep-web information acquisition. Empirical results on challenging deep IS benchmarks demonstrate that NestBrowse offers clear benefits in practice. Further in-depth analyses underscore its efficiency and flexibility.

Nested Browser-Use Learning for Agentic Information Seeking

TL;DR

NestBrowse tackles the gap between IS agents' API-level tool use and full browser-based information gathering. It introduces a minimally complete four-action browser toolkit and a nested outer-loop/inner-loop framework, trained via multi-task imitation learning to jointly optimize reasoning and intra-page exploration. On four challenging deep IS benchmarks across English and Chinese, NestBrowse—especially the 30B-A3B model—achieves strong results and outperforms many open-source baselines while rivaling several proprietary systems, with the 4B model also competitive. The approach demonstrates that principled browser abstractions and hierarchical reasoning enable small models to perform complex deep web information seeking efficiently, reducing context waste by injecting only goal-relevant content.

Abstract

Information-seeking (IS) agents have achieved strong performance across a range of wide and deep search tasks, yet their tool use remains largely restricted to API-level snippet retrieval and URL-based page fetching, limiting access to the richer information available through real browsing. While full browser interaction could unlock deeper capabilities, its fine-grained control and verbose page content returns introduce substantial complexity for ReAct-style function-calling agents. To bridge this gap, we propose Nested Browser-Use Learning (NestBrowse), which introduces a minimal and complete browser-action framework that decouples interaction control from page exploration through a nested structure. This design simplifies agentic reasoning while enabling effective deep-web information acquisition. Empirical results on challenging deep IS benchmarks demonstrate that NestBrowse offers clear benefits in practice. Further in-depth analyses underscore its efficiency and flexibility.
Paper Structure (27 sections, 8 equations, 5 figures, 3 tables)

This paper contains 27 sections, 8 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Tool-Integrated Reasoning.
  4. Deep Information Seeking.
  5. Nested Browser-Use Learning
  6. Minimally Complete Browser Toolkit
  7. Nested Browser-Use Framework
  8. Why Nested Browser-Use Matters.
  9. How Nested Browser-Use Works.
  10. Multi-Task Imitation Learning
  11. Task and Supervision Sources.
  12. Quality Filtering via Rejection Sampling.
  13. Multi-Task Imitation Learning Objective.
  14. Experiments
  15. Setup
  16. ...and 12 more sections

Figures (5)

  • Figure 1: Performance of the proposed NestBrowse on the challenging deep IS benchmark BrowseComp.
  • Figure 2: Overview of the Nested Browser-Use Framework. The outer loop interleaves reasoning and tool calls to solve the user task. Page-transition actions trigger an inner loop for intra-page exploration, which extracts and returns goal-relevant content to the outer loop, forming a nested interaction structure.
  • Figure 3: Average context length over tool-call turns for NestBrowse-30B-A3B on the BrowseComp subset. Gray bars denote the number of remaining active trajectories at each turn.
  • Figure 4: Intra-page exploration performance over inner-loop instances from 100 trajectories per benchmark.
  • Figure 5: A case study where an online browser-based calculator is used for numerical computation.