Reproduce “WebCPM: Interactive Web Search for Chinese Long-form Question Answering” 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
Yujia Qin, Zihan Cai, Dian Jin, Lan Yan, Shihao Liang, Kunlun Zhu, Yankai Lin, Xu Han and 7 more
Researchers from Tsinghua University, Renmin University, ModelBest Inc., and Tencent Inc. introduced WEBCPM, a public dataset and framework for interactive web search in Chinese Long-form Question Answering. The work provides 5,500 high-quality question-answer pairs with recorded human web search actions and a modular pipeline that demonstrates the advantages of interactive information retrieval and robust information synthesis.
Key insights
- Scaling laws hold for Chinese PLMs, with larger models (e.g., CPM-10B) consistently improving performance across interactive web search sub-tasks.
- Training the synthesis model with irrelevant 'noise' facts significantly enhances its robustness and ability to ignore noisy information, improving overall answer quality.
- Key components of the search model state, such as past actions for action prediction and existing facts for fact extraction, are crucial for mimicking human-like web search behaviors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.