AutoWebGLM: Bootstrap And Reinforce A Large Language Model-based Web Navigating Agent

• DOI: 10.48550/arxiv.2404.03648
• Publication Date: 2024-04-04
• Authors: Hanyu Lai; X Liu; Iat Long Iong; Shuntian Yao; Yuxuan Chen; Pengbo Shen; Hao Yu; Hanchen Zhang; Xiaohan Zhang; Yuxiao Dong; Jie Tang
• Keywords: FOS: Computer and information sciences; Computation and Language (cs.CL)

Large language models (LLMs) have fueled many intelligent agent tasks, such as web navigation -- but most existing agents perform far from satisfying in real-world webpages due to three factors: (1) the versatility of actions on webpages, (2) HTML text exceeding model processing capacity, and (3) complexity decision-making open-domain nature web. In light challenge, we develop AutoWebGLM, a GPT-4-outperforming automated built upon ChatGLM3-6B. Inspired by human browsing patterns, design an simplification algorithm represent preserving vital information succinctly. We employ hybrid human-AI method build data for curriculum training. Then, bootstrap reinforcement learning rejection sampling further facilitate webpage comprehension, browser operations, efficient task decomposition itself. For testing, establish bilingual benchmark AutoWebBench tasks. evaluate AutoWebGLM across diverse benchmarks, revealing its improvements also underlying challenges tackle real environments. Related code, model, will be released at \url{https://github.com/THUDM/AutoWebGLM}.