Pluralistic Alignment for Healthcare: A Role-Driven Framework

Jiayou Zhong, Anudeex Shetty, Chao Jia, Xuanrui Lin, Usman NaseemEMNLP 2025

TL;DR

Large Language Models (LLMs) need to understand diverse human values, especially in sensitive areas like healthcare. This paper introduces ETHOSAGENTS, a novel framework that dynamically simulates different personas to generate pluralistic responses, outperforming existing methods in capturing diverse ethical perspectives.

What Problem Does This Paper Address?

As LLMs become more integrated into healthcare, ensuring their outputs reflect the wide range of human values and perspectives is crucial. Current alignment techniques, including pluralistic approaches, often fall short in healthcare because personal, cultural, and situational factors heavily influence ethical viewpoints. Existing methods can lead to a homogenized understanding of preferences, overlooking critical diversity. This paper aims to bridge this gap by developing a lightweight, generalizable framework for pluralistic alignment specifically tailored for the complexities of healthcare.

Methods

The core innovation of ETHOSAGENTS lies in its dynamic persona generation and role-driven simulation. Instead of relying on static, pre-trained community LLMs (as in prior work like Modular Pluralism), ETHOSAGENTS constructs personas on the fly for each scenario. These personas are defined by six structured attributes: Name, Core Value, Ethical Framework, Right/Duty, Emotion, and Stakeholder Role. This allows for a more nuanced and adaptable representation of diverse perspectives.

The framework operates in two stages:

  1. Persona Generation: Structured prompts are used to generate multiple distinct personas for a given scenario, ensuring attribute-level diversity.
  2. Perspective Generation: These generated personas then act as conditioning inputs for the main LLM, guiding it to produce responses that reflect each persona's worldview.

This approach is model-agnostic and does not require architecture-specific fine-tuning, making it flexible and scalable. The framework supports three pluralistic alignment modes: Overton (freeform reasoning), Steerable (conditioned generation), and Distributional (population-aligned outputs).

Datasets

The primary dataset used for evaluation is VITAL (Shetty et al., 2025). VITAL was developed to address the lack of specific alignment resources in healthcare and contains 13,601 value-laden situations and 5,245 multiple-choice questions drawn from moral dilemmas, health surveys, and public opinion polls. It supports all three modes of pluralistic alignment (Overton, Steerable, and Distributional).

Key Insights and Ideas

  • Dynamic Personas Outperform Static Models: ETHOSAGENTS's ability to dynamically generate personas tailored to specific scenarios leads to better pluralistic alignment than methods relying on pre-trained, static community LLMs. This highlights the importance of context-specific perspective simulation.
  • Healthcare Demands Nuanced Alignment: The complexity of ethical decision-making in healthcare necessitates approaches that can capture subtle differences in values, rights, and duties, which ETHOSAGENTS effectively addresses.
  • Lightweight and Generalizable: The framework's ability to work without extensive fine-tuning or specialized datasets makes it particularly valuable for high-stakes, low-resource domains.
  • Attribute Importance: The study shows that a comprehensive set of persona attributes (Name, Core Value, Ethical Framework, Right/Duty, Emotion, Stakeholder Role) is crucial for generating rich and diverse ethical commentary.

Key Results

ETHOSAGENTS consistently outperforms baseline methods (Vanilla, MoE, ModPlural) across all three alignment modes on the VITAL benchmark.

  • Overton Mode: ETHOSAGENTS achieved the highest value coverage scores across multiple LLMs, demonstrating its ability to generate diverse and representative responses. For example, it achieved 44.27% value coverage with Qwen2.5-7B.
  • Steerable Mode: The framework achieved the highest accuracy in selecting the most relevant persona for alignment, with scores like 66.51% for Qwen2.5-14B.
  • Distributional Mode: ETHOSAGENTS achieved the lowest Jensen-Shannon (JS) distance, indicating a closer match to empirical human distributions, suggesting better modeling of complex moral landscapes.
  • Generalization: When compared directly to ModPlural on a shared test set, ETHOSAGENTS showed significant improvements, including a 35% relative improvement in value coverage for the Overton mode.

Limitations and Future Work

  • Language Dependency: The current study is limited to English-language inputs and outputs, restricting its global applicability. Future work should extend the framework to multilingualism to support more inclusive and regionally aware systems.
  • Inference Time: While the framework is lightweight, there is an overhead for generating personas on the fly. The authors suggest pre-computing personas and re-using them to improve efficiency.
  • Domain Specificity: While the framework is evaluated extensively for healthcare, further investigation into its generalizability for other critical domains is needed.
  • Potential for Misuse: The authors acknowledge the risk that pluralistic outputs could be misused to rationalize harmful behavior and emphasize the need for responsible use.

Summary updated

BibTeX
@inproceedings{zhong2025pluralistic,
  author = {Jiayou Zhong and Anudeex Shetty and Chao Jia and Xuanrui Lin and Usman Naseem},
  title = {{Pluralistic Alignment for Healthcare: A Role-Driven Framework}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure