PaperlayerView exact revision on arXiv

Source-linked AI summary

MedConstBench: Benchmarking How Constitutions Steer LLMs in Medical Ethics

Yibo Yang, Zerui Huang, Haomiao Yang, Zixia Wang

arXiv:2610.36513v1cs.AIcs.CL

TL;DR

Medical-ethics evaluation of LLMs has focused less on value trade-offs and often treats ethics as an intrinsic model property, leaving deployment constitutions understudied. MedConstBench evaluates constitutions with human-calibrated multi-model peer review and finds effects vary across ethical dimensions and models. It proposes a do-no-harm selector that chooses a constitution per model while matching the per-model oracle and maintaining a worst-case floor.

  • Problem

    Medical-ethics evaluations often focus on knowledge correctness or treat ethical behavior as intrinsic to the model, leaving the effect of deployment constitutions on ethical dimensions understudied.

    Source: §I · p. 1

  • Method

    MedConstBench evaluates medical-ethics responses under different constitutions using Beauchamp’s four principles and human-calibrated multi-model peer review.

    Source: §I · p. 1, A. Problem Setup · p. 3

  • Results

    Constitution effects are dimension-selective and model-dependent, and no constitution is universally optimal.

    Source: §I · p. 2, §V · p. 7

  • Takeaways & Limitations

    A do-no-harm selector chooses constitutions per model, matching the per-model oracle while guaranteeing a worst-case floor and generalizing to unseen models at low cost.

    Source: §I · p. 2

Abstract

from arXiv · show

As large language models advance rapidly, they are increasingly used to answer medical questions, and many such models are now deployed locally. Once a model is deployed, its behavior is governed mainly by the ethical policy written into its constitution. However, most existing work focuses on factual correctness rather than the trade-off between competing ethical values. Even benchmarks that do evaluate ethics treat it as a fixed property of the model rather than as a variable, so how a constitution reshapes medical-ethics behavior remains largely unexplored. We introduce MedConstBench, a medical-ethics benchmark that treats the deployment constitution as a controlled variable, is organized around Beauchamp's four principles, and is scored by human-calibrated multi-model peer review. It includes MedDilemma, our own dataset of clinical dilemmas, and supports a systematic analysis of how constitutions steer LLMs. Evaluating ten open-source LLMs under four constitutions, we find that the effect of a constitution is dimension-selective and model-dependent, and that no constitution is universally optimal. Motivated by these findings, we propose a do-no-harm adaptive selector that chooses a constitution per model, matching the per-model oracle while guaranteeing a worst-case floor and generalizing to unseen models at low cost.

I. INTRODUCTION

Medical-ethics behavior involves trade-offs among values, but existing evaluations largely emphasize factual correctness or treat ethics as intrinsic to the model. MedConstBench makes deployment constitution a controlled variable and finds constitution effects vary by ethical dimension and model, motivating per-model selection.

  • Medical questions often require trade-offs among competing values, so ethical judgments matter alongside factual correctness.Source: §I · p. 1
  • Prior medical-ethics evaluations often treat ethics as intrinsic to the model and rely on a single LLM judge.The paper argues that single-judge preferences make conclusions hard to trust.Source: §I · p. 1
  • MedConstBench evaluates deployment constitutions as a controlled variable using clinical dilemmas and human-calibrated multi-model peer review across Beauchamp’s four principles.The benchmark also evaluates on MedEthicsQA; its peer-review scores correlate with human judgment better than a single judge.Source: §I · pp. 1–2
  • Constitution effects are dimension-selective and model-dependent, with no universally optimal choice and a risk that a constitution helpful to most models can harm an individual model.Source: §I · pp. 1–2
  • The do-no-harm selector chooses a constitution per model, matching the per-model oracle while guaranteeing a worst-case floor and generalizing to unseen models at low cost.Source: §I · p. 2

II. RELATED WORK

Medical-task benchmarks commonly assess clinical knowledge or open-ended question answering, while medical-ethics evaluation is less common and varies in focus. MedConstBench addresses the gap of treating constitution as a controlled variable and using human-calibrated multi-model peer review.

A. Benchmarks for Medical Tasks

B. Constitutions.

The paper studies deployment-time constitutions as system-prompt policies that steer model behavior without changing model weights. Its problem setup evaluates locally deployable models on medical-ethics items under four constitutions and scores responses against twelve criteria.

B. Constitutions.

  • A constitution is a system-prompt ethical policy for steering deployed behavior without changing model weights.The study compares AMA, WHO, and Mitigation policies, which emphasize different orientations.Source: B. Constitutions · pp. 2–3

A. Problem Setup

  • The setup evaluates locally deployable models on medical-ethics items, each posing an ethically charged question or scenario.Source: A. Problem Setup · p. 3
  • Each item is presented to each model under four constitutions, including C0, the no-constitution baseline.The constitutions are selected as representative policies with different emphases.Source: A. Problem Setup · p. 3
  • A response is the answer of model m to item q under constitution c.Source: A. Problem Setup · p. 3
  • Responses are assessed against autonomy, non-maleficence, beneficence, and justice through twelve binary-graded criteria scored on a 0–2 ordinal scale.Source: A. Problem Setup · p. 3

B. Scoring via Multi-Model Peer Review

MedConstBench scores responses through multi-model peer review, excluding the evaluated model from its judge panel. Criterion scores are averaged across judges, within principles, and across items to produce principle and overall scores.

B. Scoring via Multi-Model Peer Review

C. Do-No-Harm Adaptive Constitution Selection

The adaptive selector chooses a constitution for each model subject to a do-no-harm constraint. It then maximizes a deployer-selected priority principle while ensuring no principle falls more than the tolerated amount below baseline.

C. Do-No-Harm Adaptive Constitution Selection

A. Experimental Setup

The experiments evaluate open-source models under four constitutions on two ethics datasets, with identical prompts and decoding settings across models. The evaluation uses a leave-one-out judge panel and yields over 3.4 million criterion-level judgments.

A. Experimental Setup

  • Ten open-source LLMs are evaluated, with three additional models held out from selection-rule derivation to test generalization to unseen models.All models receive identical prompts and decoding settings so behavioral differences are attributable to the constitution rather than prompting.Source: A. Experimental Setup · p. 4
  • MedDilemma contains 400 human-filtered clinical dilemmas, and 400 sampled MedEthicsQA items cover ethics discourse alongside clinical decision-making.MedDilemma items are generated from clinical-ethics references and filtered for plausibility, principle conflict, and quality.Source: A. Experimental Setup · p. 4
  • Figure 2 shows per-principle changes relative to the no-constitution baseline under AMA, WHO, and Mitigation on both datasets; blue and red indicate increases and decreases.Source: Figure 2 · p. 4
  • Across ten models, four constitutions, and two datasets, leave-one-out judging produces over 3.4 million criterion-level judgments.Each response is scored by nine judges on twelve criteria.Source: A. Experimental Setup · p. 4

B. Main Results

Constitutional effects vary across ethical principles: Beneficence is near-saturated, while Autonomy and Justice shift more, so no constitution improves every principle at once.

B. Main Results

  • No constitution improves all four principles at once because their responses to constitutions differ.Beneficence is near-saturated, while Autonomy and Justice are more responsive.Source: B. Main Results · pp. 4–5
  • Beneficence averages 1.89 on MedDilemma and barely changes under any constitution, while Autonomy and Justice start lower and move most.Source: B. Main Results · p. 4
  • The radar charts show Beneficence fully extended, with constitutions differing mainly along Autonomy and Justice.Source: B. Main Results · p. 5

C. Do-No-Harm Adaptive Selection

The do-no-harm selector filters constitutions against a baseline floor, then chooses by deployment priority. It matches the oracle on reported worst-case performance, generalizes to held-out models, and can use a small probe.

C. Do-No-Harm Adaptive Selection

D. Human Validation

Multi-model peer review aligns more closely with human ratings than a single judge, while judges show substantial agreement with one another.

D. Human Validation

V. CONCLUSION

MedConstBench treats deployment constitutions as a controlled variable and evaluates them with human-calibrated peer review. Its findings motivate per-model adaptive selection because effects are dimension-selective and model-dependent, with no universally optimal constitution.

V. CONCLUSION

  • Across ten open-source LLMs and two datasets, constitutional effects are dimension-selective and model-dependent, motivating a do-no-harm selector that matches the per-model oracle and generalizes to unseen models.The selector guarantees a worst-case floor and deploys at low cost.Source: §V · p. 7

Related papers

Loading 2610.36513v1…