CLExEval: A Human-in-the-Loop Framework for Qualitative Evaluation of LLM Clinical Reasoning

The most underreported finding here is not the verbosity bias itself but what it implies about how current benchmarks are constructed: most clinical LLM evaluations use information-rich prompts that inadvertently hand models the answer, meaning published accuracy scores may reflect prompt engineering quality more than genuine diagnostic reasoning.

CLExEval belongs to a growing cluster of papers exposing the gap between benchmark performance and real-world reliability. The 'Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues' paper from the same date makes a structurally identical argument in a different domain: models appear aligned or accurate under evaluation conditions, then degrade when those conditions shift even slightly. Both papers are essentially arguing that current evaluation design is the problem, not just the models. The calibration paper from June 30 ('Calibration, Not Compilation') adds a third data point: syntactic or surface correctness, whether in code or clinical language, does not guarantee functional correctness. Taken together, these suggest a broader methodological crisis in how AI outputs are validated before deployment.

Watch whether any major clinical AI vendor (Epic, Microsoft Health, Google Health) publicly commits to physician-annotated evaluation protocols within the next six months. If none do, CLExEval risks becoming a cited-but-ignored benchmark, the same fate as earlier clinical NLP evaluation frameworks that preceded it.