Reproduce “WebCoT: Enhancing Web Agent Reasoning by Reconstructing Chain-of-Thought in Reflection, Branching, and Rollback” 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
Minda Hu, Tianqing Fang, Jianshu Zhang, Junyu Ma, Zhisong Zhang, Jingyan Zhou, Hongming Zhang, Haitao Mi and 2 more
WEBCOT, developed by Tencent AI Lab and CUHK, enhances web agent reasoning in Large Language Models by reconstructing and distilling specific human-like reasoning patterns—reflection, branching, and rollback—via supervised fine-tuning. This framework achieves up to a 264.9% relative accuracy improvement on web navigation benchmarks compared to baseline models and surpasses GPT-4O on WebVoyager and SimpleQA.
Key insights
- Explicitly defining and distilling web-specific reasoning patterns (reflection, branching, rollback) through SFT significantly enhances web agent performance beyond generic LLM reasoning or costly RL.
- Verbalizing complex reasoning processes into refined, human-like Chain-of-Thought rationales is more effective for SFT than relying on the agent's original self-generated thoughts.
- A cumulative training strategy that introduces more complex reasoning only for tasks that failed at simpler stages prevents 'overthinking' and improves the agent's overall learning effectiveness and efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.