Pith. sign in

REVIEW 2 major objections 5 minor 132 references

Theory-grounded two-step prompting plus self-distillation lifts multilingual moral decisions without human labels.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 03:31 UTC pith:W5I5XK2C

load-bearing objection Solid multilingual moral-reasoning package: culturally adapted benchmark, expert grounds, and self-distillation that lifts F1 and native-language use; the proxy for “theory engagement” is weaker than claimed but the empirical package still holds. the 2 major comments →

arxiv 2607.11736 v1 pith:W5I5XK2C submitted 2026-07-13 cs.CL

MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning

Ayoung Lee , Ryan Kwon , Yunxiang Zhang , Yuxuan Liu , Peter Railton , Lu Wang This is my paper
classification cs.CL
keywords multilingual moral reasoningtheory-grounded promptingcultural adaptationself-distillationmoral foundationsnative-language reasoningMCLASH
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Language models are being asked to make moral judgments for people whose languages and cultures differ from the English-centric data that most models and benchmarks were built on. Direct translation of dilemmas erases local institutions, customs, and value systems; existing reasoning scaffolds stay static and English-only; and training for moral alignment usually needs expensive human or stronger-model supervision. This paper answers all three gaps at once. It releases MCLASH, a culturally adapted multilingual benchmark of high-stakes scenarios. It supplies MET, a two-step method that first selects situation- and culture-specific theoretical grounds drawn from psychology and philosophy, then reasons over them in the user’s native language. It further adds MET-D, a self-distillation stage that uses automatically verifiable labels from synthetic character-paired dilemmas so no external supervision is required. Across three model families the trained method raises decision quality, sharply increases native-language reasoning, and shows that the most useful grounds track real cultural patterns rather than a universal English template.

Core claim

A two-step procedure—selecting expert-curated moral-theory grounds that fit the situation and culture, then reasoning over them in the native language—combined with rejection-sampling self-distillation on synthetic character-paired dilemmas, consistently improves multilingual moral decision accuracy and native-language legibility without any external human or stronger-model labels.

What carries the argument

MET / MET-D: the model first selects a situation- and language-specific subset of 37 expert-curated theoretical grounds (organized into six dimensions: Normative Authority, Value Systems, Ethical Theories, Cognitive Reasoning Strategies, Conflict Handling Strategies, Moral Uncertainty), then generates a theory-guided reasoning chain; MET-D further trains the second step by retaining only those chains whose final ternary answer matches the automatically derived ground-truth label of a synthetic CLASH-style character.

Load-bearing premise

Matching the automatically derived yes/no/ambiguous label on synthetic character-paired dilemmas is treated as a reliable signal that the model has substantively engaged with and reconciled the selected theoretical grounds.

What would settle it

Measure whether models that reach the correct ternary label after MET-D training actually reconcile conflicting grounds (for example by human annotation of reconciliation quality or by controlled ablation of grounds) rather than merely pattern-matching the final answer; if correct labels rise while reconciliation stays flat, the proxy fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Moral-reasoning systems can be made culture-aware without translating English dilemmas or importing Western scaffolds wholesale.
  • Self-distillation on character-paired synthetic data can replace expensive human or stronger-model supervision for ternary moral decisions.
  • Beneficial moral-theory grounds differ systematically by language and track real cultural patterns (e.g., contractarianism for Confucian languages, relation-based authority for Spanish).
  • Cross-lingual transfer of moral reasoning follows linguistic typology more than cultural similarity, so same-script or same-word-order languages help each other more than culturally close but typologically distant ones.
  • Enforcing native-language reasoning becomes performance-positive rather than a trade-off once the model has been trained with MET-D.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the label-proxy assumption holds only for clear-cut cases, the same pipeline may still under-train genuine multi-framework reconciliation needed for high-stakes ambiguous dilemmas.
  • The same select-then-reason-plus-self-distill recipe could be ported to other value-laden domains (law, medicine, content moderation) that currently rely on English-centric rubrics.
  • Ground-selection itself remains untrained; a second distillation stage that rewards culturally coherent ground sets could close the remaining gap the authors leave open.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper addresses multilingual moral decision-making with three contributions: (1) MCLASH, a culturally adapted (not merely translated) benchmark of 1,852 long-form scenarios across six languages derived from CLASH; (2) MET, a two-step prompting method that first selects from 37 expert-curated theoretical grounds organized into six dimensions, then produces theory-guided reasoning in the user’s native language; and (3) MET-D, which adds rejection-sampling self-distillation on synthetic character-paired dilemmas whose ternary labels are automatically determined by constructed value priorities, requiring no external supervision. On Qwen3-4B/8B and Gemma3-4B, MET-D raises macro-F1 by ~3.71 points on MCLASH and ~4.23 on MMoralExceptQA (peak +12.94 Malay on Qwen3-8B), increases native-language reasoning by ~62 points, and yields analyses of culture-specific beneficial grounds and typology-driven cross-lingual transfer.

Significance. If the results hold, the work supplies a reusable culturally adapted evaluation resource, an extensible theory-ground inventory, and a supervision-light training recipe that improves both accuracy and legibility of multilingual moral reasoning. Strengths include multi-model/multi-dataset evaluation (MCLASH, MMoralExceptQA, MoCa), ablations separating selection from reasoning (Table 2), random-selection baselines (Table 3), utilization annotations, and explicit cultural analyses (Figures 4–6). The self-distillation design that avoids stronger-model or human labels is practically valuable. These contributions are of clear interest to multilingual NLP and AI ethics, even if absolute gains remain modest.

major comments (2)
  1. §4.2 and the axiom that matching the synthetic ternary label is a proxy for “substantive engagement with the grounds” is load-bearing for the claim of improved theory-guided reasoning. The paper’s own evidence undercuts this: §6.2 reports utilization rising only from ~52% to ~61% (English) and explicitly notes that engagement ≠ reconciliation; Figure 18 shows the trained model listing more grounds without weighing them. Without a direct measure of reconciliation quality (or human preference over reasoning chains), the F1 gains (Table 1) and native-language gains establish better label prediction under a theory-flavored prompt more securely than improved theory-guided moral reasoning. A targeted human evaluation of reconciliation, or an ablation that removes ground names while keeping character cues, would strengthen the causal claim.
  2. Training and primary evaluation both rely on the same character–value structure (CLASH-style synthetic data for MET-D; MCLASH for main results). Although MMoralExceptQA and MoCa provide external checks, the largest and most culture-specific claims rest on MCLASH. The paper should quantify how much of the MCLASH gain is explained by distributional similarity to the synthetic training set versus genuine transfer of theory-guided reasoning, e.g., by reporting performance stratified by topic/value overlap or by training only on non-overlapping topics.
minor comments (5)
  1. Table 1 and Appendix C: report statistical significance or confidence intervals for the modest average gains (~3–4 F1); several per-language deltas are small enough that variance could matter.
  2. §6.1: Cohen’s κ of 0.259 across models for ground selection is low; clarify whether this still supports “systematic, non-random” selection or mainly within-model stability (κ=0.468).
  3. Figure 18 and §6.2: the distinction between engagement and reconciliation is important; consider elevating a short quantitative reconciliation metric (even on the 20-instance sample) into the main text.
  4. Limitations correctly flags untrained ground selection and missing reconciliation; these should be cross-referenced earlier when interpreting MET-D gains so readers do not over-read the mechanism.
  5. Minor presentation: ensure consistent spelling of “Probabilistic” vs “Probablistic” in grounds lists; expand the brief MoCa drop explanation for Gemma3-4B (clear-cut bias) with a short confusion-matrix note.

Circularity Check

1 steps flagged

No load-bearing circular derivation; only mild train/test structural self-reference on CLASH-style data, with independent external benchmarks.

specific steps
  1. other [Section 4.2 (MET-D training); evaluation on MCLASH]
    "We sidestep this by constructing a synthetic dataset that follows the structure of CLASH: each dilemma is paired with a character whose value priorities make the correct decision automatically verifiable. MET-D (MET-Distillation) then applies rejection-sampling self-distillation on this dataset, using correct decisions as a proxy for substantive engagement with the grounds."

    Training labels and the primary MCLASH test set share the same character-value → ternary-label construction. Rejection sampling therefore optimizes for a signal that is definitionally available on CLASH-style data. This is mild structural self-reference, not a fitted-parameter-as-prediction or self-definitional theorem; external MMoralExceptQA/MoCa results and independent utilization/native-language metrics keep the central claims non-circular.

full rationale

This is an empirical methods paper (prompting + self-distillation + multilingual benchmark), not a first-principles or uniqueness derivation. MET-D’s rejection sampling retains chains whose final ternary answer matches synthetic character-derived labels (Section 4.2); that is standard self-distillation with a proxy objective, not a claim that a fitted parameter or self-defined quantity is a novel prediction. Evaluation includes held-out MMoralExceptQA and MoCa, where gains transfer (avg +4.23 and +3.53 F1 on Qwen3-4B), so the central performance claims are not forced by construction from the training labels. Native-language rate and per-ground utilization analyses are independent of the rejection-sampling objective. Overlap of authors with CLASH (Lee et al., 2025) is normal incremental work and is not used as a uniqueness theorem or load-bearing external fact. The skeptic’s concern—that correct labels can be reached without reconciling grounds—is a validity/proxy weakness (acknowledged by the paper in §6.2 and Figure 18), not circularity of the derivation chain. Score 1 only for mild structural self-reference: synthetic training and MCLASH both inherit CLASH’s character-paired ternary structure.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

The central performance claims rest on (1) treating language as a workable proxy for culture, (2) the expert-curated ground inventory being sufficiently complete and correctly translated, (3) synthetic character-value dilemmas supplying a faithful training signal for theory engagement, and (4) a handful of design choices (number of grounds, rejection-sampling ratio, prefill tokens) that are fixed rather than derived. No continuous free parameters are fitted to the evaluation metrics themselves.

free parameters (3)
  • number of synthetic scenarios per language = 500 / 2500
    Fixed at 500 scenarios / 2 500 characters; chosen by authors, not derived.
  • yes:no:ambiguous balancing ratio after rejection sampling = 1:1:1
    Hard-coded 1:1:1; affects which reasoning chains enter the distillation set.
  • native-language prefill tokens = language-specific strings
    Hand-chosen tokens (e.g., “Okay,” “Primero”) used to force reasoning language; not learned.
axioms (4)
  • domain assumption Language is an adequate proxy for culture when adapting moral dilemmas.
    Stated in §3 and footnote 2; underpins the entire MCLASH construction.
  • ad hoc to paper Matching the automatically derived ternary label on character-paired synthetic data is a valid proxy for substantive engagement with selected theoretical grounds.
    Core justification for MET-D rejection sampling (§4.2).
  • domain assumption The 37 expert-curated grounds spanning six dimensions are sufficiently comprehensive for culture-aware moral reasoning.
    Curated with a domain expert (Railton); treated as fixed inventory throughout.
  • domain assumption Native-speaker inspection with the provided checklist yields culturally faithful adaptations.
    Quality control for MCLASH; no quantitative IAA reported.
invented entities (3)
  • MCLASH benchmark independent evidence
    purpose: Culturally adapted multilingual evaluation set of 1 852 situations / 9 260 characters.
    New dataset constructed for this paper; independent of prior English CLASH.
  • 37 theory-based grounds in six dimensions no independent evidence
    purpose: Selectable moral-reasoning scaffolds for MET.
    Expert-curated inventory introduced here; descriptions and translations are paper-specific.
  • MET / MET-D pipeline no independent evidence
    purpose: Two-step select-then-reason prompting plus self-distillation training.
    Methodological contribution of the paper.

pith-pipeline@v1.1.0-grok45 · 32877 in / 3139 out tokens · 28872 ms · 2026-07-14T03:31:09.700137+00:00 · methodology

0 comments
read the original abstract

Language models are increasingly used for moral decision-making across diverse linguistic and cultural contexts, yet existing work overlooks multilinguality on three aspects: 1) multilingual evaluation benchmarks use direct translation, failing to adapt culture-specific items; 2) inference-time methods for moral reasoning rely on static, English-centric scaffolds and lack grounding in moral theory; 3) training methods for moral decision-making typically require expensive supervision from stronger models or human annotators. We address these gaps with three contributions. First, we introduce MCLASH, a multilingual moral decision-making benchmark to capture culturally situated moral intuitions and social norms across languages. Second, we propose MET (Multilingual Ethics with Theory-grounded reasoning), a two-step prompting method built on expert-curated, theory-based grounds drawn from psychology and philosophy: the model first selects situation- and culture-specific grounds, then reasons over them in the native language of the user. Third, we introduce MET-D (MET-Distillation), which enhances the second step through a self-distillation training stage that requires no external supervision. MET-D improves macro-F1 over the base model on all three models of different sizes and families (Qwen3-4B, Qwen3-8B, Gemma3-4B), by an average of 3.71 points on MCLASH and 4.23 on MMoralExceptQA, with a peak MCLASH gain of 12.94 points for Malay on Qwen3-8B. We further reveal that MET-D increases native-language reasoning by 62.13 points on average, and that beneficial grounds differ systematically across cultures. Together, these contributions open the path for culture-aligned, theory-grounded multilingual moral reasoning.

Figures

Figures reproduced from arXiv: 2607.11736 by Ayoung Lee, Lu Wang, Peter Railton, Ryan Kwon, Yunxiang Zhang, Yuxuan Liu.

Figure 1
Figure 1. Figure 1: Example ground selection in different languages: Grounds are organized into dimensions and paired with descriptions . For the same situation, the selected grounds vary by language: in English, First-principles reasoning reflects the Western rationalism, whereas in Korean, Contractarianism reflects Confucian values. Western cultural context into account. We address this with MCLASH, a multilingual moral dec… view at source ↗
Figure 2
Figure 2. Figure 2: Sparse theory-guided reasoning behavior in the base model. On 26-scug (En￾glish), the base explicitly mentions only two of its ten grounds, Institution-Based and Responsibility-Based , both weighing the duty to uphold ethics standards, and collapses to a definitive yes (punish) rather than the correct ambiguous, under-weighting the character’s equal commitment to forgiving an honest miscommunication. from … view at source ↗
Figure 3
Figure 3. Figure 3: Method: In MET, the model first selects situation- and language-specific grounds, then reasons based on them. MET-D extends this by adding a self-distillation training stage to improve theory-guided reasoning. The training dataset for MET-D is synthetically generated using Qwen3-32B, following a situation-and-character-description structure where ground￾truth labels are automatically derivable without manu… view at source ↗
Figure 4
Figure 4. Figure 4: Gain per dimension and ground: All dimensions yield positive gains on average, with language-specific nuances. Hindi benefits most, likely because the base model has less prior knowledge of moral reasoning in low-resource languages [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Native-language reasoning: MET-D achieves the highest average F1 and native￾language percentage in reasoning chains, with the only exceptions being Hindi F1 (a low￾resource language) and Chinese F1 and native percentage (where the base model already reasons natively without enforcement). Together, these results show that beneficial grounds are not language-agnostic: they track real cultural patterns, under… view at source ↗
Figure 6
Figure 6. Figure 6: Cross-lingual transfer: Linguistic similarity improves transfer. For some high￾resource languages (English, Chinese), train￾ing on the language itself hurts performance, while training on a different language helps. Even without target-language training, cross-lingual transfer can still help. We train Qwen3-4B on each language individ￾ually and evaluate on all six ( [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Figure for model selection descriptions, we found that Qwen3-32B already performs well in adaptation, with only minor language issues, such as occasionally generating text in languages irrelevant to the target language. Therefore, we corrected these minor issues manually to ensure linguistic quality, without applying an extensive checklist as was done for the situations. Because adaptation revises entities… view at source ↗
Figure 8
Figure 8. Figure 8: Prompt used for culturally adapting the situation, action, and reasons 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt used for culturally adapting the character descriptions 19 [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: General guidelines for annotating situation, action, and reasons Dataset Language # Situations Avg # Tokens Situation Action Char Desc CLASH (en) English 345 659.14 7.34 87.84 MCLASHes Spanish 319 787.30 11.98 141.52 MCLASHhi Hindi 277 2826.07 39.77 535.47 MCLASHkr Korean 317 1141.53 18.80 281.78 MCLASHms Malay 330 991.70 13.37 184.82 MCLASHzh-CN Chinese 264 811.36 14.80 256.19 MCLASH 6 languages 1852 116… view at source ↗
Figure 11
Figure 11. Figure 11: Task-specific checklist for annotating situation, action, and reasons 21 [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Examples of task-specific checklist items from [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Dimensions, grounds, and descriptions: We present six dimensions, each with a set of specific grounds and descriptions that the target model can select from. B.2 Prompts Used for Generating the Training Dataset The prompt used for extracting values from the WVS questionaire is presented in [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Prompt for selecting the grounds: We translate this prompt into the target languages using Google Translate to produce multilingual global set of grounds. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Prompt for selecting the grounds, continued from [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Instruction for following the grounds: This instruction is included in each benchmark’s original prompt to make the target model follow the selected grounds. It is translated into the target language using Google Translate to elicit multilingual responses. I will give you a text, and you should output values that are closely related to this text. Values are defined as: The intrinsic goods or ideals that p… view at source ↗
Figure 17
Figure 17. Figure 17: Prompt used for extracting values from WVS questionaire Method Macro F1 (%) / Native Language (%) Avg ∆ en es hi ko ms zh Base (not enforced) 71.01 65.65 52.27 56.68 62.48 55.87 58.59 — native % — 16.74 21.00 12.78 15.96 86.78 30.65 Base 70.73 64.05 45.66 51.24 63.42 57.74 56.42 −2.17 native % — 88.39 99.33 99.21 89.35 84.90 92.24 MET-D (not enforced) 74.57 65.14 49.19 57.41 63.86 63.49 59.82 +1.23 native… view at source ↗
Figure 18
Figure 18. Figure 18: Case study without reconciliation on the English instance 854-ama. Among the selected ten grounds, the base substantively engages with three ( Institution-Based , Virtue Ethics , Context-Sensitive ); MET-D engages with eight, adding Right-Based , Contractarianism , Stakeholder , Harm–Benefit , and Principle-Based . Yet broader en￾gagement does not translate into reconciliation. MET-D issues a one-line ver… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

132 extracted references · 20 linked inside Pith

  1. [1]

    Attention is All you Need , url =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , url =

  2. [2]

    Aligning AI With Shared Human Values , author=

  3. [3]

    arXiv preprint arXiv:2110.07574 , year=

    Can machines learn morality? the delphi experiment , author=. arXiv preprint arXiv:2110.07574 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2410.02683 , year=

    DailyDilemmas: Revealing Value Preferences of LLMs with Quandaries of Daily Life , author=. arXiv preprint arXiv:2410.02683 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2402.05070 , year=

    A roadmap to pluralistic alignment , author=. arXiv preprint arXiv:2402.05070 , year=

    Pith/arXiv arXiv

  6. [6]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Modular pluralism: Pluralistic alignment via multi-llm collaboration , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  7. [7]

    arXiv preprint arXiv:2504.10823 , year=

    Clash: Evaluating language models on judging high-stakes dilemmas from multiple perspectives , author=. arXiv preprint arXiv:2504.10823 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2509.21443 , year=

    One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral Reasoning , author=. arXiv preprint arXiv:2509.21443 , year=

    arXiv

  9. [9]

    Language Model Alignment in Multilingual Trolley Problems , author=

  10. [10]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Are Language Models Consequentialist or Deontological Moral Reasoners? , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  11. [11]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Exploring the choice behavior of large language models , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  12. [12]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    The staircase of ethics: Probing llm value priorities through multi-step induction to complex moral dilemmas , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    Evaluating the moral beliefs encoded in llms , author=. Advances in Neural Information Processing Systems , volume=

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Moca: Measuring human-language model alignment on causal and moral judgment tasks , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , volume=

    Civics: Building a dataset for examining culturally-informed values in large language models , author=. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , volume=

  16. [16]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Scruples: A corpus of community ethical judgments on 32,000 real-life anecdotes , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  17. [17]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Structured moral reasoning in language models: A value-grounded evaluation framework , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  18. [18]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Investigating cultural alignment of large language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  19. [19]

    Advances in Neural Information Processing Systems , year=

    When to Make Exceptions: Exploring Language Models as Accounts of Human Moral Judgment , author=. Advances in Neural Information Processing Systems , year=

  20. [20]

    arXiv preprint arXiv:2510.16380 , year=

    MoReBench: Evaluating Procedural and Pluralistic Moral Reasoning in Language Models, More than Outcomes , author=. arXiv preprint arXiv:2510.16380 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2511.10507 , year=

    Advancedif: Rubric-based benchmarking and reinforcement learning for advancing llm instruction following , author=. arXiv preprint arXiv:2511.10507 , year=

    arXiv

  22. [22]

    arXiv preprint arXiv:2505.08775 , year=

    Healthbench: Evaluating large language models towards improved human health , author=. arXiv preprint arXiv:2505.08775 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2509.21500 , year=

    Chasing the Tail: Effective Rubric-based Reward Modeling for Large Language Model Post-Training , author=. arXiv preprint arXiv:2509.21500 , year=

    arXiv

  24. [24]

    arXiv preprint arXiv:2510.07284 , year=

    Online rubrics elicitation from pairwise comparisons , author=. arXiv preprint arXiv:2510.07284 , year=

    arXiv

  25. [25]

    arXiv preprint arXiv:2511.19399 , year=

    Dr tulu: Reinforcement learning with evolving rubrics for deep research , author=. arXiv preprint arXiv:2511.19399 , year=

    Pith/arXiv arXiv

  26. [26]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  27. [27]

    arXiv preprint arXiv:2510.02263 , year=

    RLAD: Training LLMs to Discover Abstractions for Solving Reasoning Problems , author=. arXiv preprint arXiv:2510.02263 , year=

    arXiv

  28. [28]

    arXiv preprint arXiv:2503.09501 , year=

    Rema: Learning to meta-think for llms with multi-agent reinforcement learning , author=. arXiv preprint arXiv:2503.09501 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2503.09572 , year=

    Plan-and-act: Improving planning of agents for long-horizon tasks , author=. arXiv preprint arXiv:2503.09572 , year=

    Pith/arXiv arXiv

  30. [30]

    Situation ethics: The new morality , author=

  31. [31]

    2004 , publisher=

    Ethics without principles , author=. 2004 , publisher=

    2004

  32. [32]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Culfit: A fine-grained cultural-aware llm training paradigm via multilingual critique data synthesis , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  33. [33]

    Advances in Neural Information Processing Systems , volume=

    Culturellm: Incorporating cultural differences into large language models , author=. Advances in Neural Information Processing Systems , volume=

  34. [34]

    Advances in Neural Information Processing Systems , volume=

    Culturepark: Boosting cross-cultural understanding in large language models , author=. Advances in Neural Information Processing Systems , volume=

  35. [35]

    arXiv preprint arXiv:2406.14805 , year=

    How well do llms represent values across cultures? empirical analysis of llm responses based on hofstede cultural dimensions , author=. arXiv preprint arXiv:2406.14805 , year=

    Pith/arXiv arXiv

  36. [36]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Disentangling language and culture for evaluating multilingual large language models , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  37. [37]

    Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Having beer after prayer? measuring cultural bias in large language models , author=. Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  38. [38]

    Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Do moral judgment and reasoning capability of LLMs change with language? A study using the multilingual defining issues test , author=. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  39. [39]

    arXiv preprint arXiv:2404.10199 , year=

    Culture-gen: Revealing global cultural perception in language models through natural language prompting , author=. arXiv preprint arXiv:2404.10199 , year=

    Pith/arXiv arXiv

  40. [40]

    arXiv preprint arXiv:2505.15257 , year=

    When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners , author=. arXiv preprint arXiv:2505.15257 , year=

    arXiv

  41. [41]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    ShifCon: Enhancing non-dominant language capabilities with a shift-based multilingual contrastive framework , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  42. [42]

    arXiv preprint arXiv:2507.05418 , year=

    Learn globally, speak locally: Bridging the gaps in multilingual reasoning , author=. arXiv preprint arXiv:2507.05418 , year=

    arXiv

  43. [43]

    Advances in Neural Information Processing Systems , volume=

    Blend: A benchmark for llms on everyday knowledge in diverse cultures and languages , author=. Advances in Neural Information Processing Systems , volume=

  44. [44]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Question translation training for better multilingual reasoning , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  45. [45]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    MAPO: Advancing multilingual reasoning through multilingual-alignment-as-preference optimization , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  46. [46]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Langbridge: Multilingual reasoning without multilingual supervision , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  47. [47]

    Self-pluralising culture alignment for large language models , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  48. [48]

    Language Models are Alignable Decision-Makers: Dataset and Application to the Medical Triage Domain

    Hu, Brian and Ray, Bill and Leung, Alice and Summerville, Amy and Joy, David and Funk, Christopher and Basharat, Arslan. Language Models are Alignable Decision-Makers: Dataset and Application to the Medical Triage Domain. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Techn...

    doi:10.18653/v1/2024.naacl-industry.18 2024

  49. [49]

    Nature , volume=

    Large language models encode clinical knowledge , author=. Nature , volume=. 2023 , publisher=

    2023

  50. [50]

    arXiv preprint arXiv:2302.05729 , year=

    A brief report on lawgpt 1.0: A virtual legal assistant based on gpt-3 , author=. arXiv preprint arXiv:2302.05729 , year=

    Pith/arXiv arXiv

  51. [51]

    AI Open , volume=

    Lawformer: A pre-trained language model for chinese legal long documents , author=. AI Open , volume=. 2021 , publisher=

    2021

  52. [52]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Are rules meant to be broken? understanding multilingual moral reasoning as a computational pipeline with unimoral , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  53. [53]

    arXiv preprint arXiv:1808.07074 , year=

    The what, the why, and the how of artificial explanations in automated decision-making , author=. arXiv preprint arXiv:1808.07074 , year=

    Pith/arXiv arXiv

  54. [54]

    2018 IEEE 5th International Conference on data science and advanced analytics (DSAA) , pages=

    Explaining explanations: An overview of interpretability of machine learning , author=. 2018 IEEE 5th International Conference on data science and advanced analytics (DSAA) , pages=. 2018 , organization=

    2018

  55. [55]

    Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , volume=

    Decoding Multilingual Moral Preferences: Unveiling LLM's Biases Through the Moral Machine Experiment , author=. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , volume=

  56. [56]

    Proceedings of the 2025 ACM conference on fairness, accountability, and transparency , pages=

    Normative evaluation of large language models with everyday moral dilemmas , author=. Proceedings of the 2025 ACM conference on fairness, accountability, and transparency , pages=

    2025

  57. [57]

    arXiv preprint arXiv:2601.08430 , year=

    RubricHub: A Comprehensive and Highly Discriminative Rubric Dataset via Automated Coarse-to-Fine Generation , author=. arXiv preprint arXiv:2601.08430 , year=

    arXiv

  58. [58]

    arXiv preprint arXiv:2510.07743 , year=

    Openrubrics: Towards scalable synthetic rubric generation for reward modeling and llm alignment , author=. arXiv preprint arXiv:2510.07743 , year=

    arXiv

  59. [59]

    arXiv preprint arXiv:2511.16660 , year=

    Cognitive foundations for reasoning and their manifestation in llms , author=. arXiv preprint arXiv:2511.16660 , year=

    arXiv

  60. [60]

    Assessment & Evaluation in Higher Education , volume=

    Assessment rubrics: towards clearer and more replicable design, research and practice , author=. Assessment & Evaluation in Higher Education , volume=. 2017 , publisher=

    2017

  61. [61]

    1988 , publisher=

    A textbook of translation , author=. 1988 , publisher=

    1988

  62. [62]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  63. [63]

    The philosophical review , volume=

    Moral realism , author=. The philosophical review , volume=. 1986 , publisher=

    1986

  64. [64]

    2008 , publisher=

    Kantian ethics , author=. 2008 , publisher=

    2008

  65. [65]

    Theories of international relations , volume=

    Constructivism , author=. Theories of international relations , volume=. 2005 , publisher=

    2005

  66. [66]

    The American Review of Public Administration , volume=

    Social contract theory: Implications for professional ethics , author=. The American Review of Public Administration , volume=. 2006 , publisher=

    2006

  67. [67]

    Canadian Journal of philosophy , volume=

    Rawlsian constructivism in moral theory , author=. Canadian Journal of philosophy , volume=. 1987 , publisher=

    1987

  68. [68]

    Contractualism , author=

  69. [69]

    2023-02-17 , url =

    WVS , title =. 2023-02-17 , url =

    2023

  70. [70]

    PNAS nexus , volume=

    What values best distinguish the world's cultures? The machine learning-based cultural values inventory , author=. PNAS nexus , volume=. 2025 , publisher=

    2025

  71. [71]

    The psychology of legitimacy: Emerging perspectives on ideology, justice, and intergroup relations , volume=

    Theories of legitimacy , author=. The psychology of legitimacy: Emerging perspectives on ideology, justice, and intergroup relations , volume=

  72. [72]

    A companion to philosophy of law and legal theory , pages=

    Legal positivism , author=. A companion to philosophy of law and legal theory , pages=. 2010 , publisher=

    2010

  73. [73]

    Political theory , volume=

    The communitarian critique of liberalism , author=. Political theory , volume=. 1990 , publisher=

    1990

  74. [74]

    Existentialism , author=

  75. [75]

    Journal of Educational and Psychological Consultation , volume=

    Constructivism , author=. Journal of Educational and Psychological Consultation , volume=. 1993 , publisher=

    1993

  76. [76]

    Philosophy & social criticism , volume=

    Discourse ethics and civil society , author=. Philosophy & social criticism , volume=. 1988 , publisher=

    1988

  77. [77]

    Advances in experimental social psychology , volume=

    Moral foundations theory: The pragmatic validity of moral pluralism , author=. Advances in experimental social psychology , volume=. 2013 , publisher=

    2013

  78. [78]

    Online readings in Psychology and Culture , volume=

    An overview of the Schwartz theory of basic values , author=. Online readings in Psychology and Culture , volume=

  79. [79]

    Journal of cross-cultural psychology , volume=

    Hofstede's culture dimensions: An independent validation using Rokeach's value survey , author=. Journal of cross-cultural psychology , volume=. 1984 , publisher=

    1984

  80. [80]

    The nature of human values

    Rokeach value survey , author=. The nature of human values. , year=

Showing first 80 references.