EthosAgents: Pluralistic Healthcare Alignment

• EthosAgents is a pluralistic alignment framework for healthcare that dynamically constructs multiple persona-based perspectives to reflect diverse stakeholder values.
• It operates in two stages—persona generation and perspective generation—and supports Overton, Steerable, and Distributional alignment modes.
• Empirical results show significant gains such as a +36.6% improvement in value coverage on Qwen2.5-7B, highlighting enhanced ethical diversity over standard methods.

EthosAgents is a role-driven, model-agnostic, training-free pluralistic alignment framework proposed for healthcare, where acceptable outputs depend on diverse values, stakeholder positions, and situational constraints rather than on a single averaged preference. It dynamically generates multiple structured personas for each scenario, elicits persona-grounded comments, and then uses those comments differently across Overton, Steerable, and Distributional alignment modes. In the paper’s framing, the method is a “first lightweight, generalizable, pluralistic alignment approach” for simulating diverse perspectives and values in high-stakes settings (Zhong et al., 12 Sep 2025).

1. Concept and motivation

EthosAgents is motivated by the claim that healthcare pluralism is unusually difficult. The paper argues that standard alignment methods such as RLHF often optimize toward a homogenized or averaged human preference, while healthcare decisions depend on personal values, cultural and religious beliefs, stakeholder position, situational specifics, emotional framing, and ethical frameworks. A model aligned to average preference may therefore suppress minority views, flatten morally contested questions, or misrepresent genuine disagreement (Zhong et al., 12 Sep 2025).

The framework is also positioned against prior pluralistic alignment methods, especially Modular Pluralism (ModPlural). The paper argues that static community models are too coarse for healthcare, that fine-tuning community-specific models is expensive and data-hungry, and that healthcare pluralism is often scenario-specific rather than reducible to broad ideological buckets. EthosAgents responds by constructing perspectives on the fly rather than relying on fixed community experts (Zhong et al., 12 Sep 2025).

In this sense, EthosAgents is not a general-purpose ethics engine. It is an inference-time method for making pluralism explicit and operational in domains where values are heterogeneous and context-sensitive. Its design target is not only diversity of answers, but diversity of moral standpoints.

2. Framework architecture

EthosAgents operates in two stages: Persona Generation and Perspective Generation. For a scenario $s$, it first constructs a persona set

$$\mathcal{P}(s) = \{ p_1, p_2, \ldots, p_k \}, \quad \text{where} \quad p_i \sim P(\cdot \mid s),$$

and then generates a response from each persona’s standpoint,

$y_i \sim P(y \mid s, p_i).$

The framework is therefore scenario-conditional at both stages rather than relying on a fixed inventory of pre-trained moral submodels (Zhong et al., 12 Sep 2025).

Each persona is defined through six dimensions. The paper’s example illustrates the intended structure.

Dimension	Example
Name	Public Health Steward
Core Value	Collective Wellbeing
Ethical Framework	Utilitarianism
Right/Duty	Duty to Reduce Population Harm
Emotion	Relived
Stakeholder Role	Public Health Systems

This six-part schema is central to the method. The paper’s ablations later show that reducing personas to only Name, Core Value, and Right/Duty lowers performance, which implies that Ethical Framework, Emotion, and Stakeholder Role are functional rather than decorative attributes (Zhong et al., 12 Sep 2025).

Operationally, persona generation is guided by a prompt that asks for “6 contrasting ethical perspectives” that differ in Core values, Ethical framework, Key right/duty emphasized, Emotion, and Stakeholder role. Perspective generation then asks for a direct moral comment that includes “2–3 core moral values involved,” “Specific rights or duties relevant,” “Ethical judgment and reasoning,” and “Normative language” (Zhong et al., 12 Sep 2025).

The result is a structured plurality mechanism: EthosAgents does not merely sample stylistic variants of one answer, but attempts to instantiate multiple explicit ethical viewpoints around a single case.

3. Alignment modes and evaluative logic

EthosAgents is designed to support all three pluralistic alignment modes adopted in the paper: Overton, Steerable, and Distributional. In Overton mode, the generated persona comments are concatenated and summarized into one response that reflects the range of reasonable perspectives. In Steerable mode, the system selects the most relevant persona for a target stance or value orientation and conditions final generation on that perspective. In Distributional mode, multiple persona-conditioned response distributions are generated and then aggregated using priors, so that the output models a population distribution rather than a single answer (Zhong et al., 12 Sep 2025).

The evaluation uses a healthcare benchmark called VITAL, with 18,846 items overall: 1,649 Overton items, 15,340 Steerable items, and 1,857 Distributional items. The benchmark spans value-laden healthcare situations such as vaccination refusal, health-data sharing, assisted death, trust in doctors or scientists, and technology-mediated care (Zhong et al., 12 Sep 2025).

The paper evaluates seven backbone models: LLaMA2-7B, LLaMA2-13B, Gemma-7B, LLaMA3-8B, Qwen2.5-7B, Qwen2.5-14B, and ChatGPT (GPT-3.5-turbo). Persona generation uses DeepSeek-R1, while role-based perspective generation uses Qwen2.5-7B and DeepSeek-V3 in different settings. Baselines are Vanilla direct prompting, MoE, and ModPlural (Zhong et al., 12 Sep 2025).

For metrics, the paper uses value coverage for Overton, accuracy for Steerable, and Jensen-Shannon distance for Distributional. It also supplements automatic scoring with LLM-as-a-Judge and human pairwise evaluation on sampled moral scenarios (Zhong et al., 12 Sep 2025).

4. Empirical findings

The paper reports that EthosAgents is strongest in Overton mode, where it is best on 5 of 7 models. Its largest reported gain is on Qwen2.5-7B, where value coverage rises from 32.41 for Vanilla to 44.27, described as a +36.6% improvement. It also reports 42.73 on Qwen2.5-14B, compared with 31.29 for Vanilla. The method does not dominate every backbone-instance combination: for example, Gemma-7B and ChatGPT have stronger Vanilla Overton scores than EthosAgents in the reported table (Zhong et al., 12 Sep 2025).

In Steerable mode, the paper reports aggregate scores of 58.25 for LLaMA2-7B, 59.71 for Gemma-7B, 65.87 for Qwen2.5-7B, 60.71 for LLaMA2-13B, 66.51 for Qwen2.5-14B, and 64.78 for ChatGPT. The appendix shows a more mixed picture across subcategories, but the reported trend is that EthosAgents improves steerability especially in value-sensitive settings (Zhong et al., 12 Sep 2025).

In Distributional mode, the paper states that EthosAgents achieves the lowest JS distance among all backbone models overall, especially on poll-style questions, though the appendix shows that ModPlural often performs slightly better on the moral-scenario subset. The distributional result is therefore strong but not uniform across every sub-split (Zhong et al., 12 Sep 2025).

A concise summary of representative outcomes is given below.

Evaluation slice	Baseline	EthosAgents
Overton, Qwen2.5-7B	32.41	44.27
Overton, Qwen2.5-14B	31.29	42.73
Generalization benchmark, Overton	22.22	30.03
Generalization benchmark, Steerable	34.47	37.70
Generalization benchmark, Distributional	0.56	0.38

The paper’s ablations are also informative. Overton value coverage on Qwen2.5-7B drops from 44.27 to 36.35 when personas are reduced to partial attribute sets, and performance increases as the number of personas rises from 1 to 6. With fewer personas, scores move from 38.01 to 40.38 to 43.70, reaching 44.27 at six personas. Replacing Qwen2.5-7B with DeepSeek-V3 as the perspective generator often improves results further, including substantial gains on some ChatGPT settings (Zhong et al., 12 Sep 2025).

The paper also reports qualitative diversity gains. Compared with ModPlural, EthosAgents produces higher 2-gram and 3-gram diversity across all three alignment modes, and its case studies show more explicit articulation of values such as bodily autonomy, collective welfare, non-maleficence, care, integrity, and public responsibility (Zhong et al., 12 Sep 2025).

5. Relation to broader agent ethics research

Although EthosAgents is introduced in healthcare, it sits within a wider technical landscape concerned with role-sensitive behavior, interaction design, and dynamic social alignment. Related work on multi-party principal loyalty argues that once an agent represents a principal while speaking to a counterparty, the heuristic “help whoever you are talking to” becomes the wrong objective; the agent must protect the principal’s interests without broad over-refusal (Li et al., 29 Jun 2026). This is not the same problem as pluralistic healthcare alignment, but it points to a shared concern: ethical competence depends on who the agent is acting for and which stakeholder perspective is normatively relevant.

Work on the Human-AI-Experience (HAX) framework similarly treats trustworthiness as something built through interaction protocol, interface surface, output constraints, and a behavioral proxy that “filters, sequences, and adapts” agent outputs before they reach the user (Scibelli et al., 12 Dec 2025). This suggests that EthosAgents addresses one layer of the problem—pluralistic perspective construction—while interface-level governance and trust calibration remain separate design questions.

A further adjacent result comes from generative-society experiments showing that prompt-defined identities are “inherently unstable” under sustained interaction, and that agents can develop endogenous stances measured by Innate Value Bias (IVB), Persuasion Sensitivity, and Trust-Action Decoupling (Zhang et al., 24 Mar 2026). A plausible implication is that EthosAgents’ structured personas are best understood as a controlled alignment mechanism for deliberative tasks, not as a guarantee that long-horizon social agents will maintain stable moral identities in open multi-agent environments.

Taken together, these neighboring lines of work situate EthosAgents as a pluralistic alignment layer rather than a complete socio-technical governance stack.

6. Limitations and open directions

The paper is explicit that EthosAgents has important limitations. It is currently English-only, adds inference overhead because personas are generated on the fly, and can be misused if a user cherry-picks one persona-grounded perspective to justify a preferred conclusion. The persona generator itself is still an LLM, so it may omit communities, stereotype stakeholder roles, overrepresent familiar ethical categories, or hallucinate plausible but weakly grounded viewpoints (Zhong et al., 12 Sep 2025).

The framework also does not solve medical factuality, clinical safety, legal compliance, or triage accountability. Its contribution is pluralistic moral representation, not end-to-end clinical governance. The empirical results are strong overall but not uniform: EthosAgents does not win every Overton setting, and ModPlural remains slightly better on some Distributional moral-scenario subsets (Zhong et al., 12 Sep 2025).

For this reason, the paper presents EthosAgents most plausibly as a pluralistic deliberation aid. It is intended to surface value diversity, make stakeholder standpoints explicit, and support summarization, steering, or population-sensitive modeling in morally contested settings. Its broader significance lies in showing that pluralistic alignment can be made context-sensitive, stakeholder-aware, and normatively explicit without retraining separate community models for every domain (Zhong et al., 12 Sep 2025).