CMoralEval: Benchmarking AI Moral Reasoning

Updated 5 April 2026

CMoralEval is a comprehensive benchmark that rigorously evaluates moral reasoning and value alignment in language and multimodal models.
It employs multi-level taxonomies and culturally diverse annotations with high inter-annotator agreement for realistic moral dilemma assessments.
The framework utilizes both outcome-focused and process-based metrics to enable diagnostic comparisons and iterative improvements in AI moral competency.

CMoralEval is a designation widely adopted in the literature for comprehensive benchmarks that evaluate the moral reasoning, judgment, and value alignment capacities of LLMs and, in some extensions, multimodal models. Modern instantiations of CMoralEval exhibit strong theoretical grounding, cross-domain and cross-cultural breadth, explicit attention to annotation quality, and process-focused as well as outcome-focused evaluation protocols. Benchmarks within the CMoralEval family have become standard tools for quantitative and qualitative assessment of AI moral competence, supporting both diagnostic comparison and iterative model improvement across domains and languages (Yu et al., 2024, Ji et al., 2024, Chen et al., 29 Sep 2025, Chiu et al., 18 Oct 2025, Morlat et al., 24 Dec 2025, Mohammadi et al., 7 Oct 2025).

1. Conceptual Foundations and Design Rationale

CMoralEval benchmarks are rooted in foundational theories of moral psychology and ethical philosophy, most centrally the Moral Foundations Theory (MFT), but extend to other major frameworks: Schwartz’s Human Values, Curry’s Morality-as-Cooperation, Gert’s Common Morality, and various professional and cultural codes. The goal is to rigorously probe how LLMs (and related models) handle both surface-level and deep moral dilemmas, capturing their ability to (1) recognize morally salient situations, (2) align responses with theory-specific norms or human judgments, and (3) explicate their reasoning processes. Datasets are typically constructed to blend real-world case diversity, theoretical exhaustiveness, and scenario authenticity (Yu et al., 2024, Ji et al., 2024, Chiu et al., 18 Oct 2025).

2. Taxonomies and Annotation Schemes

Moral scenarios in CMoralEval are annotated using multi-level, often overlapping taxonomies that reflect both culture-specific and universal features.

Prominent taxonomies include:

Moral Foundations Theory (MFT): Care/Harm, Fairness/Cheating, Loyalty/Betrayal, Authority/Subversion, Sanctity/Degradation; Liberty/Oppression appears in some contemporary datasets.
Schwartz Human Values: Ten value domains ranging from Self-Direction to Universalism.
Morality-as-Cooperation: Family, Group, Reciprocity, Heroism, Deference, Fairness, Property.
Common Morality: Ten rules emphasizing harm avoidance, deception, and duty (Chen et al., 29 Sep 2025).

For benchmarks focused on Chinese LLMs, annotation includes five categories: Familial Morality, Social Morality, Professional Ethics, Internet Ethics, and Personal Morality, each underpinned by five Confucian and modern principles: Goodness, Filial Piety, Ritual, Diligence, Innovation (Yu et al., 2024). Rich quality controls are imposed, with inter-annotator agreement (Cohen’s κ) often exceeding 0.8, double annotation, and expert review to guarantee reliability.

3. Dataset Construction and Scenario Diversity

Data curation in CMoralEval benchmarks is characterized by systematic sourcing from authentic, culturally relevant materials:

Societal, Professional, Familial, and Online Contexts: Episodes from moral TV programs, legal/ethical casebooks, news archives, and academic corpora (Yu et al., 2024, Chen et al., 29 Sep 2025).
Cross-linguistic and Cross-cultural Sampling: Chinese, English, and increasingly, multilingual datasets anchored in both national and global value surveys (e.g., WVS, PEW) (Mohammadi et al., 7 Oct 2025).
Scenario Complexity: Datasets typically include explicit moral scenarios (with clear right/wrong), nuanced moral dilemmas (featuring plausible yet norm-violating distractors), and multi-label or multi-category problems to stress-test multi-dimensional reasoning (Yu et al., 2024).

Instance counts and coverage:

Benchmark	#Instances	Key Categories/Foundations
CMoralEval (ZH)	30,388	5 (Familial, Social, etc.)
MoVa	16 datasets/100k+	MFT, Schwartz, MAC, Gert
MoralBench	~60–80 (per set)	MFT (5–6)
COMETH (EN/FR)	300 (core actions)	6 (Gert rules/actions)
Moralise (MM)	2,481	13 (Turiel domains)

EMS: explicit scenarios; MDS: moral dilemmas.

CMoralEval benchmarks maintain a balance of positive and negative moral cases, with typical human accuracy on explicit scenarios exceeding 95%; dilemma scenarios are designed to reduce human accuracy (e.g. ∼85%) and thus increase difficulty for models (Yu et al., 2024).

4. Task Definitions and Evaluation Metrics

CMoralEval embraces a spectrum of tasks reflecting both recognition and generative reasoning:

Explicit Morality Classification: Models select the most/least moral options from provided alternatives.
Moral Dilemma/Preference Tasks: Choose the more/less acceptable solution; align with abstract value principles; identify subtle distinctions between nearly-plausible acts.
Multi-Label Norm Attribution: Assign one or more foundations/values per scenario, using all-at-once classification or classifier chains (Chen et al., 29 Sep 2025).
Story Understanding/Generation: Map stories to implied morals, or generate stories consistent with a given moral, with surface realism and value alignment as dual objectives (Guan et al., 2022).
Process-Focused Evaluation: Score not only the output label but also model reasoning chains, using rubric criteria to quantify identification, process, trade-off weighing, helpfulness, and avoidance of harm (Chiu et al., 18 Oct 2025).

Metrics:

Accuracy (for single-label or explicit tasks)
Macro/micro F1, AUC (for multi-label assignments)
Scenario-level scores based on weighted rubrics
Model-to-human correlation (Pearson’s r) when assessed against survey means (e.g. WVS/PEW) (Mohammadi et al., 7 Oct 2025)
Consistency and coverage (e.g., how often models align with majority human judgments, or achieve self-consistency across sampled completions)
Binary and comparative scores via mirror-scoring and direct item comparison (Ji et al., 2024)

5. Experimental Findings and Model Diagnostics

Benchmarks consistently demonstrate that current models—across parameter scales—struggle with nuanced or culturally divergent moral reasoning:

Chinese LLMs: Mean accuracy on moral dilemma scenarios hovers at ∼33% (chance), with larger models (e.g., Yi-34B) outperforming smaller ones mainly on Familial/Personal Morality. RLHF yields inconsistent benefit, particularly for sub-10B models. "Internet Ethics" is persistently weak (Yu et al., 2024).
Cross-cultural LLMs: Western cultural alignment is markedly stronger (r ∼0.82) than non-Western (r ∼0.61), revealing systematic bias in training or evaluation pipelines (Mohammadi et al., 7 Oct 2025).
Multi-label and Multi-category Scenarios: Model performance drops by 3–5% when multiple moral dimensions are invoked, indicating sensitivity to label correlation and scenario complexity (Yu et al., 2024).
All-at-once prompting: A simple all@once multi-label prompt strategy outperforms fine-tuned classifier chains on both in-domain and out-of-domain tasks, achieving macro F1 increases >+0.10 in some categories (Chen et al., 29 Sep 2025).

6. Theoretical and Practical Implications

CMoralEval benchmarks enforce explicit coverage of (a) multiple moral frameworks to reduce theory-specific bias, (b) culturally grounded annotation, and (c) high-quality, interpretable evaluation protocols. Their influence is seen in methodological innovations:

Rubric-based, process-oriented scoring exposes not only final judgments but also whether essential moral tradeoffs, acting principles, or context features are considered, supporting transparent audit trails (Chiu et al., 18 Oct 2025).
Lightweight LLM-based annotation/augmentation shows feasibility of scaling moral scenario creation with expert-in-the-loop or AI-assisted authoring, maintaining high inter-annotator agreement (Yu et al., 2024).
Application as diagnostic and alignment tools: CMoralEval enables practitioners to pinpoint foundation-specific or culturally specific weaknesses and guides corpus selection, RLHF reward shaping, and cross-lingual or cross-domain model adaptivity (Chen et al., 29 Sep 2025, Ji et al., 2024).

7. Limitations, Biases, and Future Directions

Despite major advances, several limitations persist:

Topic and Cultural Bias: Underrepresentation of non-Western and emerging moral issues (digital privacy, environmental ethics) is common (Mohammadi et al., 7 Oct 2025, Lin et al., 20 May 2025).
Scaling and Generalization: Most datasets are English- or Chinese-centric; expansion to other languages and global sources is ongoing.
Process Transparency: Even with rubric approaches, capturing all meaningful axes of deliberative moral reasoning—especially in open-domain dialog—remains an open challenge.

Future directions include development of culturally agile benchmarks, incorporation of multimodal and context-sensitive norms, robust integration of process-based scoring in safety-critical domains, and alignment with evolving human values through participatory, iterative human-in-the-loop protocols (Chiu et al., 18 Oct 2025, Morlat et al., 24 Dec 2025).

Key References:

"CMoralEval: A Moral Evaluation Benchmark for Chinese LLMs" (Yu et al., 2024)
"MoVa: Towards Generalizable Classification of Human Morals and Values" (Chen et al., 29 Sep 2025)
"MoralBench: Moral Evaluation of LLMs" (Ji et al., 2024)
"MoReBench: Evaluating Procedural and Pluralistic Moral Reasoning in LLMs, More than Outcomes" (Chiu et al., 18 Oct 2025)
"EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment ..." (Mohammadi et al., 7 Oct 2025)
"Morality is Contextual: Learning Interpretable Moral Contexts from Human Data with Probabilistic Clustering and LLMs" (Morlat et al., 24 Dec 2025)
"A Corpus for Understanding and Generating Moral Stories" (Guan et al., 2022)