Reproduce “AutoWebGLM: A Large Language Model-based Web Navigating Agent” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang and 3 more

AutoWebGLM develops an efficient web navigating agent using a 6-billion-parameter open-source LLM, ChatGLM3-6B, achieving performance competitive with or superior to much larger proprietary models like GPT-4 on complex web tasks. The system introduces a novel HTML simplification algorithm, a hybrid human-AI data construction method, and a multi-stage curriculum training approach, alongside the first bilingual web browsing benchmark, AutoWebBench.

Key insights

  • Smaller, open-source LLMs can achieve state-of-the-art performance in complex web navigation tasks when combined with sophisticated data processing and advanced training methodologies.
  • Effective HTML simplification and a comprehensive action space are crucial for enabling LLMs to interact with diverse and dynamic web environments.
  • A multi-stage training paradigm incorporating supervised fine-tuning, reinforcement learning (DPO), and self-play finetuning (RFT) is essential for developing robust and generalizable web agents capable of self-correction.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.