Reproduce “Towards Mitigating Hallucination in Large Language Models via Self-Reflection” 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
Ziwei Ji, Tiezheng Yu, Yan Xu, Nayeon Lee, Etsuko Ishii, Pascale Fung
Researchers at the Center for Artificial Intelligence Research (CAiRE) at HKUST developed an iterative self-reflection framework for Large Language Models to reduce hallucination, particularly in medical Generative Question Answering. This method, applied to various LLMs and medical datasets, demonstrated improved logical consistency and reduced factual and query inconsistencies in generated medical answers.
Key insights
- LLMs exhibit different types of hallucinations in medical GQA, including fact inconsistency, query inconsistency, and tangentiality, with varying prevalence across models.
- Instruction-based fine-tuning (e.g., MedAlpaca) is more beneficial for medical Generative Question Answering than non-instruction tuning.
- Problematic answers are often associated with lower frequencies of keywords or topics in general text distributions, suggesting LLMs may hallucinate more on less common concepts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.