Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions

Tao Yu, Zhengbo Zhang, Zhiheng Lyu, Junhao Gong, Hongzhu Yi, Xinming Wang, Yuxuan Zhou, Jiabing Yang, Ping Nie, Yan Huang, Wenhu Chen

TL;DR

This paper introduces BrowserAgent, a Playwright-based web agent that learns through direct, fine-grained browser interactions rather than static content wrappers. It adopts a lightweight two-stage training regime (SFT followed by RFT) and an explicit memory mechanism to support long-horizon reasoning, achieving robust performance on open-domain and multi-hop QA tasks with far less training data than prior approaches. By simulating human-like actions (scrolling, clicking, typing) and maintaining a structured memory of conclusions, BrowserAgent demonstrates strong generalization and scalability, notably a ~20% improvement on multi-hop benchmarks with a 7B model. The approach reduces reliance on external parsing or summarization tools and offers a practical, data-efficient pathway toward scalable, interactive web agents for complex information-seeking tasks.

Abstract

Efficiently solving real-world problems with LLMs increasingly hinges on their ability to interact with dynamic web environments and autonomously acquire external information. While recent research like Search-R1 and WebDancer demonstrates strong performance in solving web tasks, they heavily rely on additional tools to convert the interactive web environment into static text content. This is in contrast to human browsing behaviors, which involve diverse interactions with the browser, such as scrolling, clicking, and typing. In this paper, we propose BrowserAgent, a more interactive agent that solves complex tasks through human-inspired browser actions. BrowserAgent operates directly on raw web pages via Playwright through a set of predefined browser actions. We adopt a two-stage training (Supervised Fine-Tuning (SFT) and Rejection Fine-Tuning (RFT)) to improve the model's generalization abilities. Despite using significantly less training data than Search-R1, BrowserAgent achieves more competitive results across different Open-QA tasks. Additionally, we introduce an explicit memory mechanism to store key conclusions across steps, further enhancing the model's reasoning capabilities for long-horizon tasks. Notably, BrowserAgent-7B can achieve around 20\% improvement over Search-R1 on multi-hop QA tasks like HotpotQA, 2Wiki, and Bamboogle. These results indicate that BrowserAgent can serve as a more advanced framework for more interactive and scalable web agents.

BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions

TL;DR

This paper introduces BrowserAgent, a Playwright-based web agent that learns through direct, fine-grained browser interactions rather than static content wrappers. It adopts a lightweight two-stage training regime (SFT followed by RFT) and an explicit memory mechanism to support long-horizon reasoning, achieving robust performance on open-domain and multi-hop QA tasks with far less training data than prior approaches. By simulating human-like actions (scrolling, clicking, typing) and maintaining a structured memory of conclusions, BrowserAgent demonstrates strong generalization and scalability, notably a ~20% improvement on multi-hop benchmarks with a 7B model. The approach reduces reliance on external parsing or summarization tools and offers a practical, data-efficient pathway toward scalable, interactive web agents for complex information-seeking tasks.

Abstract

Efficiently solving real-world problems with LLMs increasingly hinges on their ability to interact with dynamic web environments and autonomously acquire external information. While recent research like Search-R1 and WebDancer demonstrates strong performance in solving web tasks, they heavily rely on additional tools to convert the interactive web environment into static text content. This is in contrast to human browsing behaviors, which involve diverse interactions with the browser, such as scrolling, clicking, and typing. In this paper, we propose BrowserAgent, a more interactive agent that solves complex tasks through human-inspired browser actions. BrowserAgent operates directly on raw web pages via Playwright through a set of predefined browser actions. We adopt a two-stage training (Supervised Fine-Tuning (SFT) and Rejection Fine-Tuning (RFT)) to improve the model's generalization abilities. Despite using significantly less training data than Search-R1, BrowserAgent achieves more competitive results across different Open-QA tasks. Additionally, we introduce an explicit memory mechanism to store key conclusions across steps, further enhancing the model's reasoning capabilities for long-horizon tasks. Notably, BrowserAgent-7B can achieve around 20\% improvement over Search-R1 on multi-hop QA tasks like HotpotQA, 2Wiki, and Bamboogle. These results indicate that BrowserAgent can serve as a more advanced framework for more interactive and scalable web agents.

Paper Structure

This paper contains 33 sections, 4 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. browseragent
  3. Framework Design
  4. Dataset Synthesis
  5. Environment Setting
  6. Training
  7. Experiments
  8. Benchmarks
  9. Evaluation Metrics
  10. Baselines
  11. Experimental Setup
  12. Experimental Results
  13. Ablation
  14. Increasing Memory Capacity.
  15. Multi-round observation v.s. Single-round observation
  16. ...and 18 more sections

Figures (4)

  • Figure 1: Evaluation Result of the Browseragent.
  • Figure 2: Comparison Between BrowserAgent and Traditional Deep Research Pipeline.
  • Figure 3: Overview of the Browseragent framework.
  • Figure 4: The evaluation loss curve of SFT.