Dynamic Moral Profiling (DMP)

• Dynamic Moral Profiling (DMP) is a computational approach that constructs and updates moral profiles reflecting the pluralistic and context-sensitive nature of human values.
• It uses Dirichlet-based dynamic value sampling to condition model outputs on empirical human judgment distributions, achieving a 64.3% improvement in alignment.
• DMP enhances ethical decision-making by integrating diverse value profiles, addressing biases, and reflecting nuanced human moral reasoning in complex dilemmas.

Dynamic Moral Profiling (DMP) denotes a set of computational and statistical techniques for constructing, updating, and leveraging moral profiles that reflect the pluralistic and context-sensitive nature of human value judgments. The concept is grounded in the observation that both individual and group moral stances are not static but evolve over time and in response to changing contexts, social influences, and the diversity of value priorities. DMP methods seek to dynamically align artificial agents, such as LLMs, with the distributional patterns, diversity, and reasoning processes found in human moral judgments, thus closing alignment gaps in ethically charged decision-making tasks (Russo et al., 23 Jul 2025).

1. Core Principles and Theoretical Foundations

Dynamic Moral Profiling is underpinned by several foundational theories and empirical observations:

• Pluralism in Human Morality: Empirical research demonstrates that humans do not converge on a single “correct” answer in moral dilemmas, but instead display a distribution of judgments, often buttressed by a broad, topic-dependent spectrum of values (Russo et al., 23 Jul 2025).
• Value Diversity: Human moral rationales reference a large and nuanced taxonomy of moral values. For example, a taxonomy of 60 value expressions was developed via inductive coding of over 3,000 human rationales; human value use is diffuse, with only 35.2% of value mentions concentrated in the top 10 (vs. 81.6% for LLMs) (Russo et al., 23 Jul 2025).
• Alignment Gaps: LLMs typically mirror majoritarian or high-consensus human judgments but diverge sharply as human disagreement rises, defaulting to narrow value sets or consensus answers—a phenomenon termed the “pluralistic moral gap.”

These insights motivate DMP’s design goal: to dynamically guide models to reproduce human-like pluralism and context-sensitive value diversity, rather than fixed or monolithic profiles.

2. Methodological Framework: Dirichlet-based Dynamic Value Sampling

DMP operates by constructing dynamic, data-driven value profiles from empirical human distributions and using these profiles to condition model outputs. The process involves:

• Empirical Base Measure ($G_0$): For a set of $V$ moral values, $G_0(v_k)$ is computed as the empirical frequency of each value in the full corpus of human rationales:

$$G_0(v_k) = \frac{1}{N} \sum_{i=1}^N \mathbb{I}[v_k \in \text{rationale}_i]$$

• Topic-specific Value Priors ($G_t$): Each dilemma is associated with a topic $t$, and a Dirichlet prior $G_t \sim \text{Dirichlet}(\alpha G_0)$ is defined, where $\alpha$ controls how much $G_t$ can deviate from the corpus-wide base. This enables the sampled moral profiles to reflect both global human tendencies and topic-level variation (Russo et al., 23 Jul 2025).
• Profile Sampling and Prompt Injection: For a dilemma $d_i$ and its $N_i$ human judgments, $N_i$ dynamic profiles are sampled from $G_t$, with each profile $p_j$ represented as a distribution over values (typically selecting the top-$k$). These sampled profiles are injected into LLM prompts, conditioning both the model’s final binary moral decision and its free-text rationale.

This framework formalizes DMP as a distributional alignment task and models the heterogeneity, diversity, and topic-dependency observed empirically in human reasoning.

3. Empirical Benchmarking: The Moral Dilemma Dataset and Value Alignment

The DMP methodology is benchmarked using the Moral Dilemma Dataset (MDD):

• Dataset Structure: MDD comprises 1,618 real-world moral dilemmas, each paired with approximately 32 human judgments, resulting in a total of 51,776 judgments. Each judgment is annotated with both a binary classification (acceptable/unacceptable) and a free-text rationale, which is further labeled with values from a 60-value taxonomy (Russo et al., 23 Jul 2025).
• Alignment Metrics: DMP evaluates value alignment using metrics such as the reduction in LLM-human judgment divergence, quantified as a 64.3% improvement in distributional alignment, and the diversity of values expressed, assessed via Shannon entropy and relative coverage outside the ten most frequent values (13.1% increase) (Russo et al., 23 Jul 2025).

A central finding is that standard LLM outputs align well only under high-consensus dilemmas. As human judgment diversity increases, baseline models default to consensus or popular moral values and exhibit significantly reduced alignment with the actual human value distribution. DMP overcomes this by directly sampling from and conditioning on human-derived value profiles.

4. Computational Integration and Prompting Strategies

DMP utilizes the sampled value profiles as dynamic conditioning information in prompts. This integration can be implemented as follows:

• Prompt Engineering: For each model output, the prompt is prefixed (or otherwise conditioned) with a set of sampled, weighted values from $G_t$. For example, a prompt may state: “In this dilemma, people typically prioritize $w_1$" title="" rel="nofollow" data-turbo="false" class="assistant-link">Value1, $w_2$" title="" rel="nofollow" data-turbo="false" class="assistant-link">Value2, ...” and instructs the model to generate both a binary moral decision and a rationale that draws on these values.
• One-shot/Multiple Sampling: While the paper leverages repeated sampling to construct response distributions for evaluation, a plausibly useful extension is to generate a single, profile-conditioned response each time for practical user-facing applications.

This pipeline is modular and independent of model architecture, and is designed to supply human value heterogeneity and pluralism as explicit model input.

5. Pluralism, Value Diversity, and Implications for Human-aligned Moral Guidance

The DMP approach produces model outputs that are not only closer, in a distributional sense, to human judgment distributions but also display greater value diversity in their rationales. This has direct implications for the ecological validity of machine moral guidance:

• Pluralistic Moral Guidance: By expressing a broader set of values, DMP-augmented models are less likely to default to majoritarian or implicitly Western-centric value frames, and are more adaptive to ambiguity and reasonable disagreement (Russo et al., 23 Jul 2025).
• Support for Human Decision-making: The pluralism inherent in DMP is particularly valuable for ambiguous dilemmas where consensus is low and monocultural value sets may be misleading—thus increasing the robustness and human-alignment of model-guided deliberation.
• Topic Sensitivity: Because profiles are sampled conditionally on topic (via $G_t$), DMP supports context-sensitive moral reasoning, reflecting, for example, the different values prioritized in professional versus familial dilemmas.

6. Limitations, Open Questions, and Future Directions

The DMP paradigm as instantiated in (Russo et al., 23 Jul 2025) presents several open avenues for further development:

• Single-shot vs. Distributional Sampling: Since real-world applications often involve one-shot advice rather than distributional aggregation, future research may address how to optimally select or construct the conditioning profile per instance.
• Comparative Frameworks: While DMP is empirically shown to outperform broader, theory-driven frameworks (such as Moral Foundations Theory), systematic comparison across frameworks remains an area for further investigation.
• Scalability and Deployment: Applying DMP in real-world, decentralized systems raises questions of data availability, updating value priors in dynamic populations, and maintaining efficiency at production scale.
• Implicit and Non-dilemmatic Reasoning: Extending DMP to less explicitly moral or broader judgment contexts requires adaptation of the value taxonomy and sampling regimes.
• Cultural Representativeness: The Dirichlet-sampling technique is constrained by the underlying demographic and cultural representativeness of the observed human rationales, highlighting the necessity of inclusive and updated datasets.

In summary, Dynamic Moral Profiling introduces a principled and empirically validated approach for bridging the pluralistic moral gap in AI systems, enabling models to dynamically reflect the diversity and context-sensitivity of human moral reasoning by conditioning on sampled human value profiles. This approach represents a significant advance in the construction of human-aligned, context-aware, and pluralistic moral advisory systems (Russo et al., 23 Jul 2025).