EVENT DOSSIER
The research team proposed the ConfRAG framework, aimed at addressing the hallucination problem that occurs when large language models provide uncertain answers. This framework is based on the ConfQA fine-tuning method, where the model conducts self-assessment before answering; it only outputs an answer when it is certain, otherwise it states “I’m not sure”. Experiments showed that ConfRAG reduced the hallucination rate on multi-fact judgment benchmarks from 20% to 40% to less than 5%. Additionally, this strategy reduced the number of retrieval operations by more than 30%, and in ideal cases, increased the model’s accuracy to over 95%.
本文提出 ConfRAG，一种仅在模型回答不确定时触发检索增强生成（RAG）的策略。该框架基于 ConfQA 微调方法，将多事实基准上的幻觉率从 20%-40% 降至 5% 以下。ConfQA 通过“仅自信时作答”的阻尼提示和原子事实陈述训练数据，使模型输出答案或"I am unsure"。ConfRAG 据此机制减少不必要的检索超过 30%，在理想情况下将准确率提升至 95% 以上。