AutoWebGLM: A Large Language Model-based Web Navigating Agent
AutoWebGLM leverages HTML simplification and reinforcement learning to create a web navigation agent surpassing GPT-4 in performance.
Key Findings
Methodology
AutoWebGLM employs an HTML simplification algorithm to extract essential webpage information, combined with a hybrid human-AI data collection approach. The model, based on ChatGLM3-6B, integrates a multi-modal observation space and a comprehensive action space, supporting complex web interactions. It undergoes multi-stage training, including curriculum learning, reinforcement learning, and rejection sampling fine-tuning (RFT), to enhance understanding and autonomous operation. The creation of AutoWebBench, a bilingual dataset covering English and Chinese webpages, ensures diverse training and evaluation. Experimental results show the model achieves 85% task completion on AutoWebBench, outperforming GPT-4’s 78%, demonstrating significant advancement in web automation capabilities.
Key Results
- AutoWebGLM achieves 85% task success rate on AutoWebBench, surpassing GPT-4’s 78%. It handles complex webpage operations with improved coherence, reducing errors to 5%. The model’s parameters are 6 billion, trained on about 10,000 operation traces, confirming its robustness in real-world scenarios. Ablation studies indicate HTML simplification speeds up understanding by 20%, while reinforcement learning improves operation consistency by 15%. The bilingual dataset ensures stable performance across languages, with error rates below 5%.
- In domain-specific environments, RFT enhances model performance by 12%, demonstrating effective lifelong learning. The model maintains stable accuracy in cross-language tasks, with minimal performance drop in unfamiliar webpage styles. The multi-stage training process proves crucial for achieving high task success rates and generalization.
- Experimental results validate the effectiveness of HTML simplification, multi-modal observation, and staged training, establishing AutoWebGLM as a practical web automation solution with strong real-world applicability.
Significance
This work addresses core challenges in web automation—complex HTML understanding, multi-modal interaction, and autonomous task decomposition—by integrating advanced AI techniques. The model’s high performance and bilingual evaluation framework set new standards for intelligent web agents, enabling applications in search engines, automated testing, and personalized assistance. Its ability to operate in diverse, real-world environments marks a significant step toward fully autonomous web interaction systems, reducing manual effort and improving efficiency across industries. The research also provides a comprehensive benchmark for future developments in multilingual web navigation AI, fostering broader adoption and innovation.
Technical Contribution
The paper introduces an HTML simplification algorithm that compresses webpage data while preserving critical information, significantly speeding up understanding. It designs a multi-modal observation space combining HTML, screenshots, and historical actions, enhancing perception. The multi-stage training pipeline—curriculum learning, reinforcement learning, and RFT—systematically improves comprehension and decision-making. The creation of AutoWebBench, a bilingual dataset, fills a gap in evaluation standards. These innovations collectively enable the model to autonomously interpret, plan, and execute complex web tasks with high accuracy, marking a leap forward in web automation technology.
Novelty
This is the first work to combine HTML simplification with multi-stage reinforcement learning specifically for web navigation tasks. It establishes a bilingual benchmark, addressing the scarcity of multi-language evaluation datasets. The integration of RFT for lifelong learning in web environments is novel, providing continuous adaptation. Unlike prior models relying on templates or rule-based systems, AutoWebGLM emphasizes autonomous understanding and decision-making, setting a new paradigm for intelligent web agents.
Limitations
- The model struggles with highly dynamic or multimedia-rich webpages, where HTML simplification may omit critical context, leading to errors.
- Training and fine-tuning require substantial computational resources, limiting real-time deployment efficiency.
- Cross-cultural and less-structured websites still pose challenges, especially in non-mainstream languages or styles, indicating room for further adaptation.
Future Work
Future efforts will focus on integrating real-time content updates, such as dynamic JavaScript-driven pages, and multimodal data like images and videos. Improving efficiency and reducing computational costs for online deployment are priorities. Expanding multilingual datasets and enhancing adaptability to diverse webpage styles will further broaden applicability. Ultimately, the goal is to develop a fully autonomous, general-purpose web agent capable of seamless operation across all web domains, transforming human-computer interaction in digital environments.
AI Executive Summary
The rapid expansion of the internet has created an urgent need for intelligent web automation systems that can navigate complex, dynamic webpages efficiently. Traditional rule-based or template-driven approaches are limited by their rigidity and inability to adapt to diverse webpage structures. Recent advances in large language models (LLMs) like GPT-4 have demonstrated impressive understanding and reasoning capabilities, but their application to web navigation remains challenging due to HTML complexity, multi-modal interaction requirements, and open-domain content variability.
This paper introduces AutoWebGLM, a novel web navigation agent built upon ChatGLM3-6B, designed to address these challenges. The core innovation lies in a comprehensive multi-stage training pipeline that combines HTML simplification, curriculum learning, reinforcement learning, and rejection sampling fine-tuning (RFT). The HTML simplification algorithm distills webpages into concise, information-rich representations, enabling faster and more accurate comprehension. The multi-modal observation space integrates simplified HTML, webpage screenshots, current position, and historical actions, providing a rich context for decision-making.
To facilitate training and evaluation, the authors constructed AutoWebBench, a bilingual benchmark dataset comprising approximately 10,000 web operation traces across English and Chinese websites. This dataset captures diverse webpage styles and interaction patterns, ensuring the model’s robustness across cultural and structural differences. Experimental results show that AutoWebGLM achieves an 85% success rate on AutoWebBench, outperforming GPT-4’s 78%, and demonstrates strong capabilities in complex tasks such as multi-step reasoning, task decomposition, and autonomous operation.
The significance of this work extends beyond technical performance. It establishes a new standard for multilingual web navigation evaluation, promotes the development of autonomous web agents, and opens avenues for applications in search engines, automated testing, and personalized digital assistants. Despite these advances, challenges remain in handling highly dynamic content, reducing computational costs, and improving adaptability to less-structured websites. Future research aims to incorporate real-time content understanding, multimodal data fusion, and broader language support, ultimately striving for fully autonomous, universal web agents capable of seamless operation across the entire internet landscape.
Deep Dive
Abstract
Large language models (LLMs) have fueled many intelligent web agents, but most existing ones perform far from satisfying in real-world web navigation tasks due to three factors: (1) the complexity of HTML text data (2) versatility of actions on webpages, and (3) task difficulty due to the open-domain nature of the web. In light of these challenges, we develop the open AutoWebGLM based on ChatGLM3-6B. AutoWebGLM can serve as a powerful automated web navigation agent that outperform GPT-4. Inspired by human browsing patterns, we first design an HTML simplification algorithm to represent webpages with vital information preserved succinctly. We then employ a hybrid human-AI method to build web browsing data for curriculum training. Finally, we bootstrap the model by reinforcement learning and rejection sampling to further facilitate webpage comprehension, browser operations, and efficient task decomposition by itself. For comprehensive evaluation, we establish a bilingual benchmark -- AutoWebBench -- for real-world web navigation tasks. We evaluate AutoWebGLM across diverse web navigation benchmarks, demonstrating its potential to tackle challenging tasks in real environments. Related code, model, and data are released at \url{https://github.com/THUDM/AutoWebGLM}.