Auditing frontier general-purpose large language models in biomedical tasks: reasoning gains, extraction limits, and...

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) are nearing clinical use, but their reliability and benchmark validity require thorough examination.
• Assessing LLMs' performance on real-world biomedical tasks is crucial for safe and effective deployment.

Purpose of the Study:

• To conduct a comprehensive, reproducible audit of frontier LLMs on biomedical text-mining and question-answering tasks.
• To evaluate LLM performance across diverse settings, including reasoning-intensive, extraction-oriented, and multimodal tasks.
• To assess the suitability of current benchmarks and identify potential misestimations of model capabilities.

Main Methods:

• A unified, human-centric audit of leading general-purpose LLMs.
• Utilized representative biomedical text-mining tasks and nine biomedical question-answering benchmarks.
• Included blinded expert adjudication to evaluate clinical plausibility and reasoning coherence.

Main Results:

• LLMs demonstrated consistent improvements in clinical reasoning and multimodal biomedical question-answering.
• Challenges remain in format-constrained tasks like span-level extraction and evidence-dense summarization.
• Expert review revealed that benchmark annotation errors can misestimate LLM capabilities.
• Cost-normalized analysis showed improved accuracy at lower costs for recent models.

Conclusions:

• General-purpose LLMs are approaching deployment-relevant reliability for clinical applications.
• Limitations in specific tasks necessitate hybrid architectures and human-in-the-loop systems.
• Safe and effective clinical integration requires ongoing expert oversight and evaluation.