Reproduce “RefGPT: Dialogue Generation of GPT, by GPT, and for GPT” 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
Dongjie Yang, Ruifeng Yuan, Yuantao Fan, Yifei Yang, Zili Wang, Shusen Wang, Hai Zhao
RefGPT, developed by Shanghai Jiao Tong University and collaborators, presents a framework for generating truthful and highly customizable multi-turn dialogues by leveraging external references with large language models. The method achieved 98% truthfulness, effectively mitigating hallucination, and was used to create two large-scale datasets, RefGPT-Fact and RefGPT-Code, for factual knowledge and programming scenarios respectively.
Key insights
- Leveraging external reference documents effectively constrains LLMs, significantly reducing factual errors and hallucination in multi-turn dialogue generation.
- Detailed control over various dialogue characteristics, such as turn count and utterance length, can be achieved by integrating specific instructions and template structures into LLM prompts.
- The method demonstrates robustness to variations in reference length and quality, with LLMs tending to adjust output length to match available reference information rather than generating ungrounded content.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.