Disentangled Moral Features

• Disentangled moral features are explicit, interpretable latent dimensions that break down ethical judgments into independent basis features.
• They utilize methods like linear mapping, probabilistic choice, and hierarchical Bayesian modeling to capture nuanced moral reasoning.
• Applications span automated ethical reasoning, NLP stance classification, and multimodal event extraction to enhance transparency and bias detection.

Disentangled moral features constitute the explicit, interpretable latent dimensions underlying moral reasoning and ethical decision-making, as operationalized in computational models, natural language processing systems, and empirical social science frameworks. The goal of disentanglement in this domain is to decompose observed judgments or behavioral data into a set of separate, minimally entangled basis features—such as distinct moral foundations or abstract value dimensions—enabling both mechanistic interpretability and fine-grained analysis of individual, group, or machine-led moral cognition. This approach has proven critical for transparency in automated ethical reasoning, cross-domain transferability of moral inferences, and empirical investigation of value pluralism and bias in artificial agents.

1. Formalization of Disentangled Moral Features

The extraction and operationalization of disentangled moral features typically begins by mapping raw, observable scenario features—such as individual attributes or linguistic tokens—into a structured latent space of interpretable moral dimensions. For example, in the canonical linear utility model of commonsense moral judgment (Kim et al., 2018), each scenario outcome is represented by a concrete feature vector $\theta$; this is mapped to a D-dimensional abstract moral feature space $\Lambda$ via a binary matrix $A$, yielding $F(\theta) = A\theta$.

Within this moral feature space, the utility of each alternative is computed as:

$u(\theta) = \mathbf{w}^\top F(\theta)$

where $\mathbf{w} \in \mathbb{R}^D$ assigns interpretable weights to each moral dimension (such as "old," "doctor," or "pregnant"). Net tradeoffs between scenarios, $U(\theta) = u(\theta_1) - u(\theta_0)$, drive probabilistic choices via a sigmoid transformation. Importantly, the linear structure ensures that each weight is a marginal contribution of that feature, enforcing disentanglement.

Hierarchical Bayesian extensions allow group-level norms to be learned concurrently with individual parameterizations, with the group mean $\mathbf{w}^g$ and covariance matrix $\Sigma^g$ capturing population-level regularities and the structure of within- and between-dimension moral disagreement.

2. Disentanglement in Moral Language and Social Data

Beyond controlled dilemmas, models increasingly target text and social data, where moral features must be inferred from noisy linguistic or behavioral signals. Techniques in this domain address two main challenges: entity/role disentanglement and latent dimension inference.

The morality frames framework (Roy et al., 2021) expands upon Moral Foundations Theory by lifting flat foundation polarities into structured predicates with role associations—for example, separating the "care target," "care provider," and "harm causer" within a single Care/Harm judgment. First-order logic rules and relational learning architectures (e.g., Probabilistic Soft Logic and DRaiL) are used to enforce constraints that maintain consistency between tweet-level and entity-level assignments, and to ensure that multiple entities in a text receive distinct, interpretable moral roles. Context-sensitivity is managed by conditioning on metadata such as political ideology or topic, and ablation studies confirm that such constraints are necessary to avoid conflation between raw sentiment and explicit moral role assignment.

Quantitative approaches in stance classification (Zhang et al., 2023) embed moral features as multidimensional vectors (using eMFD and FrameAxis methods), with each dimension corresponding to a refined moral foundation. Dimensionality reduction (PCA, UMAP) and clustering are used to further isolate salient, independent moral concerns.

3. Statistical, Embedding, and Optimization Methods

Information-theoretic and embedding-based methods play vital roles in ensuring that latent dimensions correspond to robust, disentangled moral factors. In noise audit studies (Mokhberian et al., 2022), entropy (measuring annotator disagreement) and silhouette coefficient (measuring latent semantic separation) serve as metrics for identifying noisy or conflated labels, directly supporting the induction of sharper, more interpretable foundation boundaries. Removal of high-entropy or low-silhouette instances yields data that enables models to converge on clearer, more distinct foundation representations.

The vec-tionary framework (Duan et al., 2023) applies nonlinear optimization to construct moral axes in embedding space. For each word in a validated dictionary, its projection onto a latent axis $m$ is optimized to minimize the squared error with crowdsourced moral ratings:

$$\hat{s}_i = \frac{\mathbf{w}_i \cdot \mathbf{m}}{||\mathbf{w}_i|| ||\mathbf{m}||} \qquad e_i = (\hat{s}_i - s_i)^2$$

The learned axis $m$ is then used for projecting unseen text, providing document-level metrics of strength, valence, and ambivalence along each foundation. This method surpasses both fixed dictionary and naive embedding averaging approaches at capturing nuanced, contextually appropriate moral features.

4. Disentanglement in Multimodal and Social Event Contexts

Morally relevant features are not confined to text, as shown in multimodal vision-language frameworks. MoralCLIP (Condez et al., 6 Jun 2025) extends contrastive learning (as in CLIP) by incorporating explicit MFT-based moral supervision into the embedding loss. Each image-caption pair is annotated with multi-label moral foundations, and the contrastive loss is augmented with a term aligning the cosine similarity of visual and textual embeddings with a Jaccard-index-based moral similarity:

$$L_\text{Total} = (1 - \lambda) L_\text{CLIP} + \lambda L_\text{Moral}$$

$$L_\text{Moral} = \frac{1}{N} \sum_i [ \textrm{sim}(v_i^e, t_i^e) - \textrm{sim}_\text{Moral}(M_{v_i}, M_{t_i}) ]^2$$

This forces the model to separate moral properties from purely semantic ones, leading to well-separated moral clusters in the joint embedding space.

In structured news event extraction (Zhang et al., 2023), event records are labeled with moral agents, patients, triggers, and MFT-aligned foundation labels, allowing models such as MOKA to represent and extract fine-grained, disentangled moral events, even when explicit moral language is absent, and to analyze ideological media framing.

5. Implications for Model Consistency, Alignment, and Robustness

Empirical analyses of LLMs across diverse settings converge on the importance—but also the fragility—of disentanglement in machine moral reasoning. Revealed preference studies (Seror, 19 Nov 2024) use a Priced Survey Methodology with utility theory to show that a subset of LLMs behave as if maximizing stable utility functions over moral dimensions, with each parameter corresponding to a disentangled moral preference (for example, the $b^i$ “ideal” answer for each foundation). Nonetheless, most models, even when consistent in aggregate, cluster near neutral stances, and the underlying dimensions are subject to individual model-specific weighting.

Recent work highlights significant inconsistencies and context sensitivity in LLMs' moral features. Moral hypocrisy analyses (Nunes et al., 17 May 2024) demonstrate that LLMs may display within-instrument consistency (e.g., regular answers to abstract questionnaires or real scenarios), yet fail to bridge abstract and concrete moral features—even when measured via the same foundation schema. Multi-preference evaluation (Jotautaite et al., 8 Apr 2025) reveals that state-of-the-art LLMs consistently prefer Care and Fairness, but their choices are highly sensitive to question framing, undermining claims of stable, robust disentanglement.

Further, advances in multidimensional assessment (Kilov et al., 16 Jun 2025) show that even high-performing LLMs often struggle to identify and isolate morally relevant features when moral noise is introduced. Only when benchmarks explicitly pre-highlight salient details do models excel at feature extraction and moral reasoning; otherwise, their sensitivity to ethical signal decay in more naturalistic settings is marked.

6. Theoretical Perspectives and Prospects for Embedding Ethics

Conceptual frameworks treat the set of disentangled moral features as a high-dimensional "moral problem space" $M$ (Waldner, 28 Sep 2025), onto which perceptions, social conditioning, and actions are mapped ($\epsilon: S \times I \times A \to M$). Human moral judgment is interpreted as a compressed, survival-biased projection $\tilde{M} = W M^*$ of this richer space, with $W$ representing evolutionary and cultural bottlenecks.

Various methods—sparse autoencoders, causal mediation, and cross-cultural corpus analysis—are proposed to empirically identify robust, interpretable moral directions in $M$, and metaethical positions (realism, relativism, constructivism, virtue ethics) are reframed as hypotheses about the structure and stability of these disentangled features.

Embedding moral features into representational substrates—e.g., via dedicated moral layers in LLM architectures—enables deep integration of ethical dimensions directly into latent spaces and facilitates empirical testing, continuous updating, and alignment via representational gradients rather than superficial post-hoc filtering.

7. Practical Applications and Future Directions

Disentangled moral features have been leveraged for multiple applied tasks: stance classification (with measurable improvements in predictive accuracy when moral features are introduced as independent vectors (Zhang et al., 2023)), moral event extraction and media analysis (Zhang et al., 2023), auditing and improving the reliability of LLM output (Seror, 19 Nov 2024), and controlling or aligning agent behavior in RL settings under moral uncertainty (Dubey et al., 17 Feb 2025). Factorial prompting protocols (Ding et al., 10 Aug 2025) further demonstrate that manipulating prompts along ethical frameworks (e.g., utilitarianism, deontology) serves as a diagnostic tool for uncovering latent alignment philosophies and maximizing explanation–answer consistency.

Nonetheless, persistent open challenges include the need for context-sensitive disentanglement robust to noise and adversarial labeling (Mokhberian et al., 2022, Rao et al., 2023), mitigating overfitting to frame-specific or surface features (Fitzgerald, 7 Jul 2024), and cross-cultural validation of foundation axes. Future research emphasizes the importance of combining theoretically principled feature construction with scalable empirical protocols, ultimately aiming at the integration of ethics as a structural property of artificial agents and cognitive models.

---

Disentangled moral features thus represent both a set of formal modeling assumptions and a research program spanning computational social science, cognitive modeling, and AI alignment. Their rigorous identification and operational use underpin interpretable, robust, and generalizable approaches to machine and human moral reasoning, with continuing impact on the transparency and trustworthiness of automated ethical systems.