Context Matching is not Reasoning: Assessing Generalized Evaluation of Generative Language Models in Clinical...

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing
• Clinical Assessment Validation

Background:

• Generative Language Models (GLMs) are increasingly evaluated using clinical Multiple-Choice Question-Answer (MCQA) benchmarks.
• Existing MCQA benchmarks, designed for human examinees, may not accurately reflect GLM capabilities due to unique model characteristics.

Purpose of the Study:

• To validate the generalizability of MCQA-based assessments for GLMs.
• To investigate the assumptions of knowledge application, semantic consistency, and null-answer recognition in GLMs.
• To identify limitations of current benchmarks and propose more robust assessment designs.

Main Methods:

• Four clinical MCQA benchmarks were validated using eight GLMs.
• Model performance was analyzed by ablating parameter size and reasoning capabilities.
• Prompt permutation was employed to test three core assumptions of MCQA generalizability.

Main Results:

• All tested GLMs demonstrated significant failures in recognizing null-answer scenarios.
• Large GLMs showed greater resilience to prompt perturbations than smaller models.
• Despite retaining knowledge, smaller GLMs exhibited a propensity for answer memorization rather than application.

Conclusions:

• Current MCQA benchmarks have questionable validity for assessing GLMs in clinical settings.
• Fundamental assumptions regarding knowledge application, semantic consistency, and null-answer recognition are globally invalidated for GLMs.
• Adaptations to benchmark design are crucial for more accurate and reliable evaluation of GLM clinical capabilities.