MedConstBench is a medical‑ethics benchmark organized around Beauchamp's four principles. It treats the deployment constitution as a controllable variable and uses human‑calibrated multi‑model peer review for scoring. The benchmark includes our own clinical‑dilemma dataset, MedDilemma, enabling systematic analysis of how constitutions steer LLM ethical behavior.
We evaluated ten open‑source LLMs under four different constitutions and found that the effect of a constitution is dimension‑selective and model‑dependent; no single constitution is universally optimal.
Motivated by these findings, we propose a “do‑no‑harm” adaptive selector that chooses the most suitable constitution per model, matching the per‑model oracle while guaranteeing a worst‑case floor, and generalizes to unseen models at low cost.
Review: This benchmark exposes the fine‑grained influence of constitutions on medical ethics, offering a practical adaptive approach for safer model deployment.