Benchmarking retrieval-augmented large language models in biomedical NLP: Application, robustness, and self-awareness

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Artificial Intelligence (AI)
• Machine Learning (ML)

Background:

• Large language models (LLMs) can generate hallucinations.
• Retrieval-augmented LLMs (RALs) mitigate hallucinations by retrieving external knowledge.
• The efficacy of RALs in biomedical NLP tasks is not well-established.

Purpose of the Study:

• To introduce a comprehensive benchmark for evaluating RALs in biomedical NLP.
• To assess RALs' performance across various tasks and robustness testbeds.
• To propose methods for enhancing RALs' robustness and negative awareness.

Main Methods:

• Developed the Biomedical Retrieval-Augmented Generation Benchmark (BARGE).
• Evaluated RALs on five biomedical NLP tasks and 11 datasets.
• Utilized four testbeds: unlabeled, counterfactual, diverse robustness, and self-awareness.
• Proposed a detect-and-correct strategy and contrastive learning for improvement.

Main Results:

• RALs generally outperform standard LLMs in biomedical NLP.
• RALs exhibit limitations in robustness and self-awareness, especially in counterfactual and diverse scenarios.
• Proposed methods significantly enhance robustness to unlabeled and counterfactual data.
• Improved models' ability to detect and avoid incorrect predictions.

Conclusions:

• Current RALs show potential but require refinement for biomedical applications.
• Robustness and self-awareness remain critical challenges for RALs in healthcare.
• Further research is needed to ensure the reliability and accuracy of RALs in high-stakes biomedical settings.