Large Language Models, scientific knowledge and factuality: A framework to streamline human expert evaluation.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Science

Background:

• Large Language Models (LLMs) offer potential for accelerating biomedical discovery by processing vast scientific literature.
• Current LLM applications face challenges in accurately inferring and extracting complex biomedical information.

Purpose of the Study:

• Introduce a novel framework to streamline the evaluation of factual scientific knowledge encoded by LLMs.
• Assess the capabilities of eleven state-of-the-art LLMs in biomedical knowledge tasks, specifically antibiotic discovery.

Main Methods:

• A three-step evaluation framework assessing fluency, alignment, coherence, factual knowledge, and specificity.
• Task delegation between non-experts and domain experts to optimize evaluation efficiency.
• Systematic assessment of LLMs on chemical compound definition and relation determination tasks.

Main Results:

• LLMs demonstrate improved fluency but exhibit low factual accuracy and bias towards over-represented entities.
• The reliability of LLMs as standalone biomedical knowledge bases is questionable.
• The need for systematic evaluation frameworks for LLM-generated biomedical knowledge is highlighted.

Conclusions:

• Current LLMs are not suitable for zero-shot biomedical factual knowledge retrieval.
• Emerging properties suggest improved factuality with domain specialization, increased model scale, and human feedback.