acceptodds
Under review as a conference paper at ICLR 2027

One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral Reasoning

Abstract

Large Language Models (LLMs) are increasingly deployed across multilingual and multicultural settings, yet it remains unclear whether changing language leads models to adopt community-specific moral reasoning or merely changes how shared learned abstractions are expressed. We conduct a controlled multilingual evaluation across six geographically, culturally, and linguistically diverse languages (Arabic, Chinese, English, Hindi, Russian, and Spanish), using parallel moral reasoning benchmarks with English-origin, Chinese-origin, and natively elicited ground-truth judgments. Across 13 open-weight LLMs spanning 2B-70B parameters, we find substantial cross-lingual divergence in moral judgments, with English generally achieving the highest performance even when ground-truth judgments originate in Chinese or are collected natively in each language. Yet the reasoning underlying these divergent judgments is considerably more convergent: Utilitarianism dominates in five of six languages, reasoning follows broadly shared stages, and language-specific moral-value associations correspond only sparsely and inconsistently to values measured in the corresponding human communities. Finally, a large-scale OLMoTrace analysis of pretraining data sources reveals little direct reproduction of training text across languages, while the corpus composition, training stage, and cultural provenance of retrieved training evidence vary substantially by response language. Thus, similar moral reasoning structures emerge even from heterogeneous and often linguistically localized training evidence. Our findings, collectively, reveal a central disconnect in multilingual moral reasoning: language changes models' moral judgments and the training evidence associated with their reasoning, but does not correspondingly localize the moral abstractions they apply.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multilingual LLMs, Moral reasoning, Cross-cultural AI

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bS2wOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines how LLMs’ moral judgments, their generated reasoning, and their alignment with human moral values vary across languages. Evaluating 13 open-weight LLMs on four moral reasoning benchmarks in six languages, the authors measure cross-lingual differences in decisions using weighted F1 and prediction disagreement. They analyze generated explanations with moral-foundation dictionaries and LLM-based annotations of ethical frameworks and reasoning stages. They relate model decisions to human moral-value profiles and use OLMoTrace to characterize lexical overlap between model outputs and training corpora. The authors find substantial cross lingual disparities in moral judgments, even though the ethical frameworks and reasoning structures are broadly similar across languages. They also find limited correspondence with community specific human values and heterogeneous training corpus matches. Together, the author claimed that shared moral abstractions may emerge without corresponding cultural localization.

Critical strengths
  1. Broad, complementary multilingual evaluation. The paper evaluates 13 models from five families across six languages and several moral reasoning benchmarks. Its mix of multilingual datasets, together with independently collected multilingual human judgments, offers a solid basis for studying cross lingual disparities beyond the usual English centric evaluations.

  2. A useful distinction between moral judgments and expressed reasoning. Rather than stopping at prediction accuracy, the paper asks whether differences in final judgments are reflected in the generated ethical explanations and moral-value associations. This multi level view is interesting and could help clarify how multilingual LLM behavior shifts across linguistic contexts.

Critical weaknesses
  1. Language dependent moral judgments are not adequately disentangled from translation and general language competence effects. RQ1 argues that cross lingual differences cannot stem from translation or source language effects because they persist on UniMoral, which the paper describes as natively authored (Section 3; Appendix A.1). However, the original UniMoral paper (Kumar and Jurgens, 2025, Section 3.1) states that its scenarios and action options were machine translated from English using SeamlessM4T; only the human judgments were collected independently in each language. UniMoral therefore does not rule out translation or source-language framing effects. More broadly, gaps in weighted F1 may reflect differences in comprehension, instruction following, or output format compliance rather than moral preferences. Appendix Table 3 even shows a case where Hindi yields no parseable prediction. Without separating these factors, the interpretation of performance gaps as moral misalignment is not well supported.

  2. The central conclusions about shared moral abstractions and cultural misalignment rest on insufficiently validated measurements. The paper treats similar generated explanations as evidence of shared moral reasoning. These explanations, however, may be prompt induced or post-hoc rationalizations rather than reflections of how the model actually reaches its decisions. This concern is sharper because several tasks explicitly invoke ethical frameworks, and the main reasoning analysis covers only a single model. It also relies on an LLM judge with limited multilingual human validation. The cross cultural validity of the measurements is also unclear. Translating and back-translating the English eMFD does not establish measurement equivalence across cultures. The predefined ethical-framework taxonomy may also miss culturally specific moral concepts, which makes the claim of predominantly Western-oriented reasoning hard to substantiate. In RQ3, merging Equality and Proportionality into a single Fairness dimension discards a distinction that is central to the cross-culturally developed MFQ-2 (Atari et al., 2023) and may mask meaningful differences. Likewise, the absence of significant associations between human MFQ-2 profiles and model decisions does not by itself indicate misalignment unless those profiles are shown to predict the corresponding human decisions. Finally, the lexical overlap measured by OLMoTrace cannot establish how shared moral abstractions form causally. As a result, the observed convergence and misalignment may partly be artifacts of the measurement framework rather than properties of the models' moral representations.

Questions for authors
  1. The original UniMoral paper says its scenarios and actions were translated from English, but this submission calls them natively authored. Could the authors explain this discrepancy and how it affects the claim that UniMoral rules out translation and source language effects?

  2. Could the authors report parsing failure rates and prediction distributions for each language? Do the main cross lingual differences hold under a simple answer only prompt that reduces differences in instruction following and output formatting?

  3. Beyond translation quality, what evidence shows that the translated eMFD measures the same constructs across languages? Why are Equality and Proportionality merged in the MFQ-2 analysis, and do the Figure 3 findings hold with the original six dimensions? As a sanity check, do these value profiles predict the human annotators' own decisions?

2 comments
AuthorsOpenReview ↗

Response to reviewer

We thank the reviewer for their review. Below we address each weakness and answer the questions asked. Changes are highlighted in blue in the manuscript.

[UniMoral “natively authored” vs SeamlessM4T] We agree the UniMoral paper states that scenarios and action options were translated from English with SeamlessM4T-v2 and manually verified, while human judgments and questionnaire profiles were collected independently in each language. When our manuscript said “natively authored,” we meant “natively annotated”. The majority-vote decisions (and associated value ratings) are produced by native speakers of each language of L, not by translating an English gold label into the target language. Because decision selection is annotated natively, there is no translation error in the evaluation labels, the target of weighted F1 and of cross-lingual label/model disagreement. That is the property that distinguishes UniMoral from MEQ/ETHICS (and from non-Chinese CMoralEval), where the gold decision is fixed in a source language and travels with a translated item. We do not claim that Seamless stimuli are free of English framing; we claim that human decision selection on those items is not a translated label, so UniMoral still functions as our native-label control for RQ1 alignment analyses. Change in manuscript: We explicitly elaborate on UniMoral’s scenario and options being translated wherever it is mentioned in this context. Specifically, in Section 3, Section 5, Appendix A.1, and Appendix A.4.

[Parse failures, prediction distributions, answer-only prompts] We agree that differences in weighted F1 may reflect not only variation in moral judgments but also differences in language comprehension, instruction following, and adherence to the required JSON output format. To help disentangle these factors, we report the mean parsing success rates across all evaluation suites: Arabic (88%), Chinese (93%), English (97%), Hindi (85%), Russian (91%), and Spanish (94%). These relatively high rates suggest that formatting failures alone are unlikely to fully explain the observed cross-lingual performance differences, although they may still contribute. We additionally report the gold-label and model prediction distributions across languages (percentages), with predictions computed over successfully parsed responses:

LanguageGold 1Gold 2Pred 1Pred 2
English54.145.951.848.2
Spanish53.146.949.750.3
Chinese59.840.251.348.7
Russian55.744.348.151.9
Arabic61.338.743.556.5
Hindi64.435.621.278.8

While predictions are relatively balanced in English, Spanish, Chinese, and Russian, we observe a stronger tendency toward Label 2 in Arabic and particularly Hindi, despite Label 1 being more prevalent in the gold distributions. This suggests that systematic prediction biases may contribute to the lower performance observed in these languages, beyond parsing failures alone. Change in manuscript: We incorporated these distributions and the parsing statistics in Section 3 and a new Appendix A.4. Finally, we agree that answer-only prompting would provide a useful control for potential differences in reasoning and formatting requirements. We will aim to complete these additional experiments before the camera-ready submission and include the results if available.

[Generated explanations and LLM-as-judge validation] Our RQ2 analysis examines the moral frameworks and reasoning phases expressed in model-generated explanations, rather than attempting to recover the models' internal decision processes. We use a predefined taxonomy of 10 ethical frameworks and 15 reasoning phases to systematically compare these explanations across languages under controlled prompting conditions. The analysis focuses on Qwen-3.5-27B, with annotations produced by Llama-3.3-70B and preliminary human agreement assessments conducted on English (n=20) and Hindi (n=20) responses. Importantly, applying the same prompts and annotation framework across languages enables a controlled comparison of expressed reasoning patterns, even though generated explanations need not faithfully reflect internal computations. Our findings therefore concern the consistency of expressed moral reasoning across languages, rather than identical underlying decision mechanisms. Similarly, our characterization of predominantly Western-oriented reasoning refers specifically to the prevalence of frameworks captured by our predefined, Western-rooted taxonomy; it does not imply that models lack culturally specific moral concepts outside this taxonomy. Change in manuscript: Clarified the scope of RQ2's analysis of expressed reasoning and the limitations of LLM-as-judge annotations, and reported preliminary human validation results in Section 4.

[Cross-lingual eMFD validity] We use translated eMFD lexicons as a common operational framework for examining associations between moral-value language and model judgments across six languages. To assess translation fidelity, we conducted back-translation checks, obtaining semantic similarity scores of 0.81 (Hindi), 0.76 (Spanish), 0.74 (Chinese), 0.78 (Russian), and 0.68 (Arabic). These checks provide an assessment of lexical preservation across translation, rather than establishing full cross-cultural measurement equivalence. Accordingly, our analysis is best interpreted as identifying exploratory associations between lexicon-based moral signals and model judgments, rather than direct comparisons of culturally invariant latent moral constructs. Change in manuscript: We made this distinction explicit in Appendix A.1.2.

[OLMoTrace and shared moral abstractions] Our OLMoTrace analysis provides complementary evidence about the relationship between generated moral reasoning and potentially relevant training-text passages. The observed verbatim span density is low (approximately 0.03), decision labels are not captured within the matched spans, and retrieved passages are predominantly English and DCLM-derived, with corpus composition varying across languages. Together, these observations suggest that the generated explanations are not predominantly attributable to direct verbatim reproduction of retrieved passages. We use this analysis as descriptive evidence about lexical overlap and retrieval patterns, rather than as a causal account of how shared moral abstractions emerge during training. Change in manuscript: Clarified in Section 6 that low verbatim overlap and language-dependent retrieval patterns characterize observable textual associations rather than establishing shared moral abstractions or causal training-data influences.

Reviewer_pNmmOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper evaluates moral judgments across six languages, thirteen open-weight LLMs, and four benchmarks. It reports substantial variation in decisions alongside greater similarity in classified reasoning frameworks, and examines associations with human moral values and retrieved training text. Specifically, the paper asks whether changing language alters LLMs' moral judgments and reasoning structures (RQ1-2), how judgments relate to moral framing and human values (RQ3), and whether training-data matches reveal reproduction or generalization of moral reasoning (RQ4).

Critical strengths
  • Broad evaluation across model families, languages, and benchmarks of different origins.

  • Useful distinction between final judgments and the ethical frameworks expressed in explanations.

  • Training-data tracing provides an interesting additional descriptive perspective.

Critical weaknesses

The paper is difficult to place; there is quite some related work to this topic. And albeit I find the more recent and broad evaluations interesting, and there is quite extensive related work discussion, to RQ1 and 3 i'm having a hard time finding the novelty and appropriate differentiation to replication. Language competence, translation effects and moreover evaluation framework biases also remain relevant to interpreting differences as moral preferences.

RQ2 and 4 i find more novel but clean evaluation and again separation to what is already there lacks behind the statements. I.p. in RQ4 remains unclear; although the authors acknowledge that OLMoTrace matches do not establish causal influence, some conclusions about shared moral abstractions and heterogeneous training-data pathways remain stronger than the evidence supports. The cross-cultural comparability of the moral-value measures and ethical-framework annotations needs stronger justification.

Questions for authors

Which findings go beyond Hämmerl et al. (2023) on multilingual moral biases, Agarwal et al. (2024) on language-dependent ethical reasoning, Khandelwal et al. (2024) on judgments versus reasoning scores, and Abdulhai et al. (2023) on moral prompting and downstream behavior and Schramowski et al. (2020) on training-data-dependent moral biases? These relevant predecessors appear absent from the references.

Does the reported framework convergence persist under neutral prompts and across additional generating models?

What evidence beyond low verbatim overlap distinguishes moral abstraction from paraphrase, translation, or recombination of training content?

2 comments
AuthorsOpenReview ↗

Response to reviewer

We thank the reviewer for their thoughtful assessment and for recognizing the potential contributions of RQ2 and RQ4. We appreciate the opportunity to clarify how our work builds on prior research and how the four research questions collectively contribute to understanding multilingual moral reasoning in LLMs. Changes are highlighted in blue in the manuscript.

[Novelty and differentiation from prior work] Our contribution is not simply demonstrating that moral judgments vary across languages, which has been established in prior work, but systematically examining where these differences emerge and how they relate across multiple analytical levels. Specifically, we connect (i) cross-lingual moral decisions across datasets with different linguistic origins (RQ1), (ii) the ethical frameworks and reasoning phases expressed in generated explanations (RQ2), (iii) associations between moral-value signals, human moral profiles, and model decisions (RQ3), and (iv) lexical overlap with retrieved training documents (RQ4). This combination allows us to investigate whether variation in final moral judgments is accompanied by corresponding variation in expressed reasoning, moral-value associations, and training-text retrieval patterns. Our contributions differ from the studies suggested by the reviewer in several respects:

Prior workOur additional contribution
Hämmerl et al. (2023)Parallel dilemma evaluations across 13 open-weight instruction models and multiple dataset origins, including language-specific human judgments
Agarwal et al. (2024)Broader model and dataset coverage, incorporating English-origin, Chinese-origin, and UniMoral evaluations
Khandelwal et al. (2024)Analysis of free-form explanations through 10 ethical frameworks and 15 reasoning phases, alongside decision-level moral-value associations
Abdulhai et al. (2023/2024)Joint examination of multilingual decisions, expressed reasoning, moral-value associations, and training-text retrieval
Schramowski et al. (2020)Multilingual analysis of verbatim training-text overlap and retrieved corpus composition using OLMoTrace
Jin et al. (2025)Broader evaluation suites, explanation-level analysis, training-text retrieval, and a taxonomy of cross-lingual failure modes

Thus, RQ1 provides a multi-origin evaluation of cross-lingual decision variation, while RQ3 examines how moral-value signals relate to those decisions, rather than repeating the performance comparison. RQ2 and RQ4 add complementary explanatory and training-text perspectives. Change in manuscript: We have strengthened the Related Work section to make these contributions and their relationship to existing findings clearer.

[Language competence, translation effects, and cross-cultural comparability] We agree that cross-lingual differences must be interpreted in light of language competence, translation effects, and evaluation-framework biases. Our experimental design addresses these concerns through complementary evaluation settings, including English-origin datasets, Chinese-origin CMoralEval, and UniMoral, which provides independently collected human judgments across six languages. These different dataset origins allow us to examine whether observed patterns are specific to a single source-language setting. However, UniMoral's scenarios were themselves machine-translated, and these comparisons do not fully eliminate translation or language-competence effects. Change in manuscript: We explicitly elaborate on UniMoral’s scenario and options being translated wherever it is mentioned in this context. Specifically, in Section 3, Section 5, Appendix A.1, and Appendix A.4.

[Framework convergence under neutral prompts and additional models] Our RQ2 analysis examines the ethical frameworks and reasoning phases expressed in Qwen-3.5-27B's generated explanations under a consistent step-by-step moral prompting protocol, with annotations provided by Llama-3.3-70B. Using a common prompting and annotation procedure enables systematic comparisons of expressed reasoning across languages while controlling for variation in the elicitation protocol. The observed similarities therefore characterize framework convergence under this experimental setting, rather than establishing invariance across prompts or generating models. We agree that neutral-prompt and multi-generator evaluations would provide valuable additional evidence about the generality of these patterns. Change in manuscript: We clarified the scope of the current analysis and identified neutral prompt and multi generator evaluations as important directions for future work in Section 4.

[OLMoTrace and evidence for shared moral abstractions] Our RQ4 analysis extends beyond measuring verbatim overlap by examining the distribution and provenance of retrieved training-text matches across languages. Specifically, we observe low verbatim span density (approximately 3%), no overlap between matched spans and parsed decision labels, and predominantly English-heavy, DCLM-dominated retrieval neighborhoods, with differences in corpus composition and training-stage distributions across response languages. We additionally examine retrieved URLs to characterize the linguistic distribution of source hosts. These findings provide descriptive evidence that model-generated moral explanations are not predominantly reproduced verbatim from the retrieved training passages, while revealing commonalities and differences in the training-text neighborhoods associated with multilingual outputs. Importantly, our analysis does not assume that low lexical overlap excludes paraphrasing, translation, or recombination, nor does retrieval establish causal influence on model decisions. Change in manuscript: Made these differences clearer in Section 6.

Reviewer_XTpaOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors assert that it’s not clear if varying language within LLMs also varies the moral reasoning to be more aligned with the culture associated with that language. Within the paper, the authors aim to contribute analysis of this behavior within 13 open-weight models by: exploring whether the models bias their moral preferences to those that within certain languages, whether models have different moral reasoning approaches in different languages, if the underlying moral theories behind different scenarios affects the responses of models, and whether or not these models simply reproduce text from their training data to generate responses.

Critical strengths

Attempts to tackle the problem from multiple approaches Evaluates a large suite of LLMs

Critical weaknesses

The authors frequently make claims or establish metrics that are not statistically valid. The authors make a number of claims that are not justified by the evidence provided (described in the “support for recommendation” section). In particular, the claims without sufficient statistical justification are foundational to the core contributions of the paper.

First, the lack of human translation validation when the authors translated ETHICS, MoralExceptQA, and CMoralEval are an important motivator behind my recommendation. The author’s attempt to justify the faithfulness of the translations to the source text, did not have high cross-lingual similarity scores, and the entirety of the paper is built upon an assumption of these relatively good translations.

Second, the invalid findings presented as answering RQ2 are a major issue in the fidelity of the data analysis. The authors cannot claim that “utilitarianism dominates in 5/6 languages” because the plot shows how different languages vary in the frameworks shown.

Third, the findings asserted to be shown from Figures 3a and 3b are a major reason. Figure 3a shows the amount of morally-salient words that are used in the text, while 3b is not related to the morally-salient words used in human decisions because UniMoral does not contain human reasoning traces of a similar form to what the models generated.

Finally, the authors use an invalid approach to quantify the moral reasoning of language models. The number of morally-salient words used in a reasoning trace does say anything about their reasoning. Each of these issues, in addition to the feedback and questions provided below, are core to the findings asserted within the paper.

It is also not clear what the novelty of this work is above: Jin, Zhijing, et al. "Language model alignment in multilingual trolley problems." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors

“OLMoTrace analysis of pretraining data sources reveals little direct reproduction of training text across languages, while the corpus composition, training stage, and cultural provenance of retrieved training evidence vary substantially by response language.” I honestly don’t know what is meant by this. Are they saying training data is typically only in one language and not multilingual?

“To what extent can LLMs generalize moral reasoning across languages and cultures? More specifically, what moral assumptions embedded in their training data limit their cross-cultural ethical competence?” It is unclear whether this is a good or a bad thing. Specifically, what does generalizing moral reasoning mean? Does it mean applying the same morals no matter the culture, or does it mean adapting to fit the culture? I’m assuming they mean the latter, but they should define what is meant by generalizable moral reasoning. Also, why are the authors asserting generalizable moral reasoning to be a desirable trait? I would assume that a model should have consistent, unshakable morals to avoid issues with safety. The authors should defend this stance, or at least clarify their motives to be exploration-based rather than argument-based.

“English dominates regardless of gold truth origin.” It is extremely unclear what is meant by this sentence. Do the authors mean that the moral values associated with English moral value datasets are most likely to be reproduced from models? If so, then stating “english dominates” does not convey this. What are the number of samples in MoralExceptQA + ETHICS versus CMoralEval and UniMoral? In the description for RQ3: “Do the moral framing of inputs in different languages and models’ inherent values” This seems like two questions. What is meant by moral framing? What are the models inherent values (are models’ inherent values these the preferences described in RQ1 or something different)? It seems like the authors are really asking about the first part, the aspect of models’ inherent values seems irrelevant to this question (though perhaps it is a justification for the results).

“retrieved evidence varies substantially” Unclear what is meant by “retrieved evidence” It is unclear what is meant by “heterogeneous and often localized training-data pathways.” Heterogeneous linguistically?

“The answers to RQ1 to RQ4 reveal systematic cross-linguistic variation in LLM moral judgment, reasoning, and value weighting, suggesting that current models rely more on shared training abstractions than on culturally grounded moral norms.” Don’t the results state the opposite? If English-related moral values are dominant and there is low verbatim overlap, doesn’t that imply that there is low cross-linguistic variation?

“LLMs exhibit substantial cross-lingual variation, yet English generally leads regardless of the source of ground-truth judgments.” The meaning of this sentence is unclear. Do they exhibit cross-lingual variation in their preferences? If so, then why does the sentence also imply that there is little variation because English leads. What is meant by “English generally leads,” is this talking about the values from English datasets? It seems like there is an equivocation of moral judgments and preferences here, when they are not the same. You should at some point define what is meant by an ethical/moral preference and a judgment. Based on my reading, you use judgment to mean a single output, while a preference is a trend across outputs (i.e., your judgments will trend toward your preferences). If you are using the definitions from UniMoral, then state that.

“Despite Chinese being the source language, it does not yield the strongest performance, and English still outperforms it.” Did the authors look into the data to understand potential reasons why? “further shows that cross-lingual behavioral similarity does not systematically reorganize around the language from which the ground truth originates.” Its unclear what is meant by “systematically reorganize.”

“we adopt an “LLM-as-a-judge” approach to automatically annotate explanations with fifteen reasoning phases and scores across ten normative ethical frameworks.” Did you verify these annotations? “We have the generated responses for all eight datasets in English (n = 26,382 traced records)” 8 datasets? There were only two discussed? If you are treating each of the dataset splits separately, then you should be consistent throughout the paper or clarify here.

“Which training corpora are associated with moral reasoning, and does their contribution differ across languages?” This doesn’t make much sense. How would the training data be specifically associated with reasoning? If the data were more likely to be produced from the model, that says nothing about its link to the models reasons and the ways that it uses that data. “Although OLMo-2 is primarily an English language model (Appendix Table 4), it successfully generates responses in all six languages evaluated here” How good are the responses? “Arabic and Spanish contain substantially larger fractions whose cultural provenance cannot be confidently identified.” How was the cultural provenance of the data identified? In the appendix it describes pulling from page metadata, but this should also be mentioned within the core paper since it is fundamental to the research question being answered.

“our results reveal shared abstraction but heterogeneous training-data pathways across languages.” What is meant by this? The authors should be very clear about the level of abstraction that is shown. Specifically, they only show the models are not exactly reproducing data from their many datasets, not that the models are actually learning to generalize from the samples provided.

“localization of retrieved training evidence does not necessarily translate into localization of the moral abstractions expressed by the model.” It is unclear what is the exact link the authors aim to make across moral abstractions and moral reasoning. In addition, you can’t make this claim because different models have different sources and different proportions of these sources. The authors have already shown in Figure 1 that models tend to perform the best when producing outputs in English, so models that were trained on a majority of English data would exhibit this pattern where the localization of retrieved training evidence does not necessarily translate into localization of the moral abstractions (unless the authors mean for “moral abstractions” to mean the literal moral reasoning text, which is not the definition of moral abstraction).

In A.1.2: “native speakers additionally verify ambiguous or low-confidence items,” Its unclear how/if you recruited native speakers. Did you just use all-MiniLM-L6-v2 instead of a native speaker?

Also in A.1.2: “The average similarity scores across languages are: Hindi (0.81), Russian (0.78), Spanish (0.76), Chinese (0.74), and Arabic (0.68).” These similarity scores do not seem great, especially not for Arabic. You cannot use these as valid translations, in addition to the many other issues I mentioned above. The authors should also seek to explain (or at least address as an open question) why the similarity scores for similarly structured languages like English and Spanish, were still less similar than very different languages like Hindi and English.

What is the authors’ approach to hierarchical clustering? Single, complete, or average. This could change a lot.

Additional issues: Combining model outputs without having a section detailing the variances in model performance/training data makes the results quite noisy. In particular, the results from RQ4 are quite inconclusive for a number of reasons, this being one.

“(Kohlberg, 1981; Haidt & Graham, 2007)” These are quite odd citations if you’re just talking about the foundations of moral reasoning. The authors should describe what they mean by moral reasoning, then use their citations to support this definition.

“moral reasoning itself is far from universal” Once again, a definition regarding what is meant by moral reasoning/the author’s stance on morality would go a long way toward supporting this statement. This seems very hand-wavey right now. For example, a moral objectivist might view moral reasoning as the attempt to approximate true morality while a relativist might treat this reasoning process as true morality in itself. Even an antirealist might need to clarify that this moral reasoning only exist as a human concept rather than a universal one.

“(Kumar & Jurgens, 2025)” This is not a good paper to cite here, the goal is to show the philosophical justification for the computational methods you will be discussing. This paper is already computational. The primary difference between the authors’ translations in comparison to UniMoral’s translations, were that the UniMoral dataset validated the outputs with human evaluators. The translations the authors created for MoralExceptQA and ETHICS are still not valid.

UniMoral is not a “native benchmark.” The data was created in English then translated to other languages and verified using human evaluators. Data of a particular origin should be pulled from that country. UniMoral would be a dataset with annotations from native annotators.

The findings for RQ2 are not consistent with the graphs shown. Specifically, the graph does not show that “Utilitarianism dominates in 5/6 languages, with Hindi uniquely favoring Deontology.” Instead, it shows models vary in the frameworks their reasoning support. In addition, this method of quantifying moral reasoning does not show whether they systematically reason in different ways, it only shows their patterns of morally-salient language use across language.

The authors should change use the term “moral foundation” or “moral theory” to clarify what they mean in different contexts, instead of “moral framing.” A large portion of the time they just mean moral foundations in MFT.

In RQ4: “To what extent does pretraining data shape LLMs’ moral orientations, and does it lead them to generalize or merely reproduce content?” This is a very big question that the data/evaluation does not support (see more below). Specifically, the claim about generalization is too broad for what the authors show.

In the evaluation for RQ4, “verbatim overlap is only ≈ 3%” is this low overlap not just explained by the model’s temperature and other forms of exploration of the generation space? How can you assert that the outputs are not semantically equivalent, because the question specifically asks how the data shapes model morals. I would even argue that semantics are not enough if you are aiming to assert full generalizability, but verbatim overlap is certainly not enough. If this is not what the authors aim to show, then they should reduce the scope.

“The answers to RQ1 to RQ4 reveal systematic cross-linguistic variation in LLM moral judgment, reasoning, and value weighting, suggesting that shared moral abstractions emerge through heterogeneous and often localized training-data pathways” These research questions don’t suggest this. Moral abstraction is not at all related to verbatim overlap, it is concerned with the moral values embedded within moral judgments/decisions. This means that RQ4 is irrelevant.

“We translate English MoralExceptQA and ETHICS, and Chinese CMoralEval using SeamlessM4T-v2 (Communication et al., 2023) in all the languages in L.” You need to validate these translations. You need to show the English F1 score on all of these datasets. Just knowing that it performs the best isn’t enough.

“These three results show that the English advantage cannot be explained solely by English-origin ground truth or translation effects. It persists when ground-truth judgments originate in Chinese” I don’t think you can assert that because the error is so great and the difference in average so small in CMoralEval. This is magnified by the fact that it is unclear how many samples are used to get these error amounts. Either emphasize statistical significance, or don’t make this claim.

In the caption for Figure 2b: “Utilitarianism dominates across languages except Hindi, as shown by item-weighted ethical-framework scores.” That is not at all what this shows? Different languages lean toward different theories.

“We obtain eMFD dictionaries for the languages in L by translating the English eMFD using a standard translation and back-translation strategy (Singh et al., 2025). Further details on dictionary translation and verification are provided in Appendix §A.1.2.” This seems like an approach that would lose a high amount of information since the dataset was only originally produced in English. Different languages have different nuances in the words that are used, and that does not get captured unless you re-annotate it. “Most languages nevertheless exhibit a sharp change at the final step, suggesting that the moral-value composition of the concluding stage differs from the preceding reasoning trajectory.” It is also still not clear how much of this error is due to poor translation/semantic equivalence. The plots shown to validate the semantics of the translations are not shown to be statistically significant (discussed below). “40 model reasonings (20 English and 20 Hindi) using the same phase and framework definitions. Reasoning-phase annotations show substantial agreement with the LLM judge.” This is quite a low number, especially if one is doing a correlation analysis (discussed below). In addition, it is never stated how many samples there are in total.

“Cohen’s κ = 0.73 for English and 0.68 for Hindi” 20 samples per language is just too noisy (even though there are 15 categories). The authors should annotate more samples, and show the proportion of categories that exist in the samples to ensure sample rarity isn’t the main cause of a high agreement. In Figure 2b, the meaning of the colors should be labelled.

“Utilitarianism emerges as the dominant framework,” Figure 2b does not show this. What it might show is that the aggregated performance across languages leans toward utilitarianism; however, it does not show that utilitarianism is a “dominant” framework, because each language individually leans toward different frameworks (with certain languages having extremely similar values for multiple frameworks). This conclusion is not valid.

“0.38-0.56 of item-weighted framework assignments” What is this metric? It is very unclear what is meant by this, and nowhere is item-weighted framework assignments defined.

“moral foundations predict permissibility in strongly language-specific ways, with only partial correspondence to established human cultural patterns.” Figures 3a and 3b are also measuring different things. Figure 3a shows the amount of morally-salient words that are used in the text, while 3b is not related to the morally-salient words used in human decisions because UniMoral does not contain human reasoning traces of a similar form to what the models generated.

In Figure 5 you should justify why you’ve chosen to not have the x and y axes be mirrored in language order

In A.4: “Chinese does not form a distinct source-language-centered cluster” Yes it does, in CMoralEval the chinese and english rows are closest then spanish is the next closest to this cluster, meaning that there is a source-language-centered cluster. In addition, these plots are simply showing the languages where the model is less likely to disagree with itself when switching language. This means that plotting results for UniMoral is invalid, because responses are different for different languages. This means that that disagreement is meaningless for that dataset. At the very least there should be a “disagreement baseline” that measures the disagreement in ground truth evaluations. Even then, it is a dubious metric. Also in A.4: “Moreover, disagreement persists on natively authored UNIMORAL, where translation is absent.” Translation is not absent in UniMoral, it is just validated by humans.

In A.5: “Across all three subsets, we observe that translations of the same scenario cluster closely together regardless of language, while remaining distinct from other scenarios.” 10 samples per group is not statistically valid. The authors should either improve how they justify their subset selection and take a statistically valid sample of each (they must provide the total number of samples in each subset), or they should plot all of the samples.

In Figure 7: It is not detailed whether the color coding is applied ad-hoc or not. If it is not applied ad hoc, then a legend should be included. Despite this, the authors should still have some way to show that the groups are actually comprised of samples that are semantically the same (the legend could serve this function). Also, describe the basis/meaning of the hierarchical clustering shown in this figure. “This indicates that the semantic content is preserved across translations” This doesn’t show what the first sentence claimed: “whether the automatic translation process influences the performance of LLM predictions across languages”

4 comments
AuthorsOpenReview ↗

Response to reviewer

We thank the reviewer for their detailed assessment and constructive suggestions. We appreciate the opportunity to clarify several methodological choices, statistical results, and interpretations that are central to our contributions. Below, we address the major concerns and identify clarifications that we have incorporated into the revised manuscript, highlighted in blue.

[Human validation of translations] We would like to clarify that our translation pipeline included human evaluation, rather than relying exclusively on automated semantic similarity metrics. Specifically, we evaluated 20 examples per language (120 language-specific examples across six languages), comparing five translation systems: Llama-3.1-8B, Llama-3.1-70B, Llama-3.2-3B, Google Translate, and SeamlessM4T-Large-2.3B (Appendix A.1). Evaluators fluent in the respective languages assessed translation quality along three dimensions: (i) content preservation, measuring whether the original meaning was retained; (ii) syntactic correctness, assessing grammatical and structural quality; and (iii) cultural appropriateness, assessing whether the intended meaning was appropriately conveyed in the target language. Based on this comparative evaluation, we selected SeamlessM4T for translating the benchmarks. Importantly, these assessments were conducted on the translated outputs themselves, rather than relying solely on back-translation or embedding similarity. We additionally examined semantic preservation using multilingual sentence embeddings (Appendix A.5), observing that translations of the same scenarios generally cluster together across languages. We use this analysis as a complementary diagnostic of semantic consistency rather than as a formal guarantee of translation equivalence. Change in manuscript: Expanded Appendix A.1 to describe the human evaluation procedure, selection criteria, and sample sizes more explicitly.

[Dataset origins and UniMoral] Our evaluation incorporates complementary sources of moral ground truth: English-origin MoralExceptQA and ETHICS, Chinese-origin CMoralEval, and UniMoral, which contains independently collected human judgments across six languages. We would like to clarify that UniMoral's scenarios and action options were translated from English, while the human moral judgments were collected independently in each language. Thus, UniMoral provides a valuable evaluation against language-specific human judgments, rather than a translation-free benchmark. These evaluation settings allow us to examine the sensitivity of observed differences to dataset origin, while recognizing that translation and general language-competence effects cannot be completely eliminated. Change in manuscript: We explicitly elaborate on UniMoral’s scenario and options being translated wherever it is mentioned in this context. Specifically, in Section 3, Section 5, Appendix A.1, and Appendix A.4.

[Clarifying the English advantage] Our finding that English generally performs best refers to higher weighted F1 against ground-truth moral judgments, suggesting greater alignment with the moral preferences represented in English-origin ground truth. However, higher F1 may also reflect differences in language competence and translation quality, rather than cultural alignment alone. We agree that performance differences should be interpreted with attention to model-level variation and statistical uncertainty. To examine this more closely, we compared English and Chinese performance on Chinese-origin CMoralEval across all 13 models. English achieves higher weighted F1 in 6 models, while Chinese performs better in 7. The mean difference (English − Chinese) is approximately 0.058, whereas the median difference is −0.010. A paired Wilcoxon signed-rank test yields p = 0.79, and a 10,000-sample bootstrap gives a 95% confidence interval of [−0.041, 0.185]. Thus, we do not observe a statistically reliable English advantage over Chinese on CMoralEval. Instead, the supported finding is that Chinese-origin ground truth does not systematically make Chinese the highest-performing response language. Change in manuscript: We revised the corresponding claims to distinguish aggregate performance patterns from statistically supported language advantages in Section 3.

[What the RQ2 analysis measures] RQ2 examines the moral concepts and reasoning structures expressed in model-generated explanations through three complementary measurements: (i) eMFD-based lexical moral-foundation signals across reasoning steps, (ii) ethical-framework scores assigned by an LLM judge, and (iii) the presence of 15 reasoning phases. These measurements characterize different aspects of expressed moral reasoning; we do not equate morally salient word frequencies with reasoning quality or assume that generated explanations directly reveal the models' internal decision mechanisms. Our primary analysis uses explanations generated by Qwen-3.5-27B, the highest-performing model in our evaluation, with Llama-3.3-70B providing framework and phase annotations. Using a common prompting and annotation protocol across languages enables systematic comparison of expressed reasoning patterns under controlled elicitation conditions. Change in manuscript: Clarified the scope of RQ2 in Section 4.

[Interpretation of ethical-framework distributions] We appreciate the opportunity to clarify Figure 2b. The figure presents item-averaged scores across ten ethical frameworks, rather than mutually exclusive assignments to a single theory. For each response, the judge produces a ten-dimensional framework-score vector, which is aggregated across items within each language. Under this operationalization, Utilitarianism receives the highest mean framework score in five of six languages (approx 0.38-0.56), while Deontology receives the highest score in Hindi (approx 0.38). This does not imply exclusive reliance on these frameworks, and the distributions also reveal substantial variation in the relative prominence of other ethical perspectives. Change in manuscript: We formally defined the aggregation procedure and clarified that the heatmap colors are column-wise min–max normalized for visualization in Section 4 and Caption of Figure 2. We also replaced the term "dominates" with the more precise statement that Utilitarianism has the highest mean annotated framework score in five languages in Section 2.

[Human validation of LLM annotations] We additionally conducted expert validation of the LLM-generated annotations using 40 explanations (20 English and 20 Hindi). Reasoning-phase annotations achieve Cohen's κ of 0.73 and 0.68 for English and Hindi, respectively, while ethical-framework scores exhibit mean absolute errors of 0.07 and 0.14 against expert annotations. These results provide preliminary evidence of correspondence between the automated annotations and expert assessments for both reasoning phases and ethical frameworks. The validation sample also includes phases with varying prevalence: Ethical Framework Application appears in 75% of human annotations in both languages, while less frequent phases such as Consultation, Implementation, and Reflection appear in at most 15%. Change in manuscript: We reported the category frequencies alongside agreement statistics and clarified the scope of the validation in Section 4.

[Distinguishing the RQ3a and RQ3b analyses] Figures 3a and 3b investigate complementary relationships between moral values and model decisions, using different sources of moral-value information. Figure 3a examines whether moral-foundation signals extracted from scenario text using eMFD are associated with model permissibility judgments. We construct five-dimensional lexical moral-foundation profiles for the scenarios and use logistic regression to estimate their associations with model decisions across languages. Figure 3b instead uses UniMoral's independently collected human MFQ-2 profiles and moral judgments. Importantly, the MFQ-2 profiles are available at the individual annotator level, rather than being inferred from generated explanations or aggregated into a single dilemma-level profile. In our analysis, we preserve each annotator's MFQ-2 profile and associate it with the model's decision on the corresponding dilemma, repeating the model decision for each associated human annotation. This allows us to examine associations between individual human moral-value profiles and model decisions without requiring human reasoning traces. The two panels therefore address distinct but complementary questions: Figure 3a characterizes the relationship between lexical moral-foundation cues in the input and model decisions, whereas Figure 3b examines the relationship between human moral-value profiles and model choices on corresponding dilemmas. Change in manuscript: We clarified this distinction and the observation-level structure of the regressions in Section 5 and Figure 3 caption.

[RQ3 MFQ-2 dimensions] We combined Equality and Proportionality into a single Fairness dimension to align the six-dimensional MFQ-2 representation with the five-foundation eMFD framework used elsewhere in the paper. This was an analytical harmonization choice, rather than an assumption that Equality and Proportionality are theoretically interchangeable. More broadly, the eMFD and MFQ-2 analyses provide operationalized associations between moral-value measurements and decisions. Change in manuscript: We explicitly clarified this rationale in Section 5.

[Scope of the RQ4 analysis and retrieval patterns across languages] Our OLMoTrace analysis examines lexical overlap between OLMo-2-32B's generated responses and passages retrieved from indexed training corpora. It provides a complementary perspective on the textual sources associated with multilingual outputs, rather than a direct reconstruction of the model's internal reasoning or a causal attribution of its moral decisions. Beyond the approximately 3% verbatim span density, we examine the composition of retrieved documents across training corpora, training stages, and response languages. For English responses, more than 90% of retrieved documents originate from DCLM, with 96.4% associated with pretraining. The retrieved corpus composition differs across non-English languages: DCLM accounts for approximately 51% of pooled Chinese retrievals and 18% of Russian retrievals, while remaining the largest source for Arabic, Hindi, and Spanish UniMoral responses (72-78%). These findings demonstrate that the distribution of retrieved training-text matches is not uniform across response languages. We use "retrieved evidence" to refer specifically to documents returned by the attribution index for matching response spans, rather than documents established to have causally influenced generation. Differences in retrieved corpus and training-stage composition therefore characterize the model's observable lexical retrieval patterns, not distinct causal learning pathways. Change in manuscript: Made these differences clearer in Section 6.

[Linguistic and cultural provenance] We additionally examine the linguistic and cultural provenance of URL-bearing retrieved documents. For example, approximately 97% of English, 89% of Chinese, and 85% of Spanish URL-bearing retrievals correspond to the respective response languages. Cultural provenance is estimated heuristically using domain information, country-code top-level domains, URL paths, and available snippets, with uncertain cases retained as such. Change in manuscript: We moved this methodological description into the main text and clarified that linguistic correspondence and inferred cultural provenance are distinct measurements in Section 6.

[Interpreting low verbatim overlap] Across English responses, approximately 3.1% of generated tokens overlap verbatim with retrieved training spans, with similarly low densities observed across other languages. Matched spans are concentrated primarily in intermediate explanations rather than final decision labels, and correct and incorrect judgments exhibit similar overlap rates. These observations provide evidence of limited direct verbatim reproduction of retrieved passages. They do not rule out paraphrasing, translation, semantic recombination, or the effects of generation temperature. Change in manuscript: We refined our RQ4 conclusions to emphasize sparse lexical copying and language-dependent retrieval composition.

[Moral judgments and preferences] A moral judgment refers to an individual model decision concerning an action's permissibility or selection, whereas a moral preference refers to a systematic tendency in judgments across multiple scenarios. Weighted F1 measures agreement with reference judgments and should not be interpreted as a direct measure of moral preferences. Change in manuscript: We clarified our operational terminology in Section 2 and standardized wordings across Section 1, Section 2, and Section 3.

[Cross-lingual generalization] Our study is descriptive rather than prescriptive. We do not assume that models should adopt different moral standards in every cultural context, nor that maintaining consistent safety-relevant principles is undesirable. Instead, we investigate how moral decisions and expressed justifications vary across languages and how these patterns relate to human judgments, moral-value measurements, and training-text retrieval. Our goal is to characterize this variation and its possible sources, rather than advocate a particular philosophical position on moral universalism or relativism.

[Novelty relative to Jin et al. (2025)] Jin et al. (2025) investigate multilingual alignment between LLM decisions and human preferences in trolley-style moral dilemmas. Our study extends this line of inquiry in several complementary directions: (i) evaluating 13 open-weight models across a broader set of moral tasks and three ground-truth origin settings; (ii) examining expressed ethical frameworks and reasoning phases alongside final judgments; (iii) relating model decisions to lexical moral-foundation signals and individual human MFQ-2 profiles; and (iv) analyzing multilingual training-text retrieval through OLMoTrace. Our contribution therefore lies not merely in documenting that moral judgments vary across languages, but in examining these differences through complementary behavioral, explanatory, value-association, and training-text perspectives. Change in manuscript: We added an explicit comparison between the two studies in Section 7.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.