acceptodds
Under review as a conference paper at ICLR 2027

Beyond Isolated Cultures: Cultural-Sphere Priors Improve LLM Alignment and Narrow the Low-Resource Gap

Abstract

Large language models often skew cultural values toward Western and English-speaking populations, while existing alignment methods treat each culture independently. We ask whether cultural-sphere structure can serve as a prior for alignment, using the Inglehart–Welzel spheres to relate cultures with similar value profiles. Across 10 models and 3 benchmarks (WVS, VSM, and open-ended PRISM), sphere-aware prompting consistently outperforms culture-only baselines, with a mean relative gain of +12.1% on WVS and particularly large gains for low-resource cultures. To identify the source of improvement, we ablate the prior and trace the gain to within-sphere consensus; the prior also recovers more human-like cultural geometry and reduces low-resource drift toward dominant spheres. We further show that fine-tuning on same-sphere consensus can internalize the sphere prior for cultures that lack training data, producing alignment that rivals an oracle trained on held-out labels and generalizes beyond the training distribution. Overall, our results support cultural-sphere relations as a useful prior for alignment across both prompting and fine-tuning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
cultural alignment, large language models, cultural sphere priors, low-resource cultures, value alignment, in-context learning, fine-tuning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jwGUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper uses cultural spheres to guide prompting and provide proxy supervision when a culture lacks training labels. Experiments across ten LLMs report better cultural alignment, particularly on WVS and for selected low-resource cultures. The fine-tuning experiments test whether supervision from related cultures transfers to held-out cultures and other benchmarks.

Critical strengths
  1. The paper addresses a practical problem: improving alignment when a target culture has no annotations. Holding out the labels of eight low-resource cultures provides a meaningful test of whether supervision can transfer from related cultures.

  2. Prompting improves WVS accuracy across all ten models, with the average rising from 52.25% to 58.51%. The ablations help explain the gain: same-sphere agreement provides most of the benefit, while distant-sphere contrast adds little.

Critical weaknesses
  1. The controls do not separate the benefit of sphere membership from that of selecting agreeing examples. The experiments lack matched controls using agreeing countries selected without sphere labels or shuffled groupings. Agreement-only prompting already reaches 58.27%, compared with 58.51% for full SC-ICL. It is therefore unclear what the taxonomy adds beyond selecting and repeating supportive examples.

  2. The VSM comparison requires a calibration that the paper does not provide. Appendix C.4 rescales model-derived scores using theoretical extrema, then compares them directly with published Hofstede scores. Putting scores in the same numerical range does not make their scales comparable; the VSM manual, §4 https://geerthofstede.com/wp-content/uploads/2016/07/Manual-VSM-2013.pdf explicitly cautions against such comparisons. This weakens the main evidence for transfer beyond WVS.

  3. The open-ended results are less convincing. The PRISM judge sees each candidate's generation prompt, including sphere information supplied only under CSP. The comparison thus changes both the response and the context used to score it, with no blinded or human validation to separate their effects. Fine-tuning also reduces PRISM performance: every checkpoint falls below its untuned baseline, with relative declines of 9.4% to 15.9% (Table 6).

Questions for authors
  1. Do the gains persist when agreeing reference countries are selected without sphere labels, with the same demonstration questions, reference counts, and prompt length?
  2. What validates the mapping from model-derived VSM scores to published Hofstede scores? Do comparisons of country rankings support the same conclusions?
  3. Does the PRISM gain persist under blinded judging with identical evaluation context for both methods?
  4. How does proxy supervision affect performance on questions where a held-out culture disagrees with its sphere's consensus?
2 comments

Thanks for your supportive review and valuable suggestions. Our responses to the weaknesses (W) and questions (Q) are provided below. A revision of our paper will be uploaded, with the revised parts marked in blue.

W1: Contribution of Cultural-Sphere Structure vs. Selected Examples

We would like to clarify the contribution of cultural-sphere priors versus the selected answers through existing and additional experimental evidence.

(1) Existing Evidence. Table 1 shows that adding cultural-sphere context without any reference answers (Vanilla → CSP) yields average relative gains of +1.5%, +3.7%, and +3.3% on WVS, VSM, and PRISM, respectively. On WVS, with the same demonstration questions and target-country answers, adding sphere context raises accuracy from 52.89% to 53.54% (Standard ICL → SP-ICL), while further incorporating sphere-based reference answers raises it to 58.51% (SC-ICL).

(2) Additional Reference Selection Ablation. To examine whether predefined spheres provide useful guidance beyond simply adding reference answers, we compared three strategies on GPT-4o, GPT-5, and GPT-5.2, keeping the demonstration questions, target-country demonstration answers, and number of references fixed.

  • Random selection samples the same number of reference countries as SC-ICL from countries with answers to each demonstration question, without using sphere membership.
  • Similarity-based selection ranks countries by the similarity of their human response distributions on separate training questions, then selects the closest countries with answers to each demonstration question, without accessing test answers.
  • Sphere-based corresponds to SC-ICL.
ModelSphere-basedRandomSimilarity-based
GPT-4o60.66%57.75%58.62%
GPT-559.52%56.72%55.41%
GPT-5.257.46%55.05%55.05%
Average59.21%56.51%56.36%

Sphere-based selection outperforms both alternatives across all three models, supporting the value of cultural relationships beyond additional reference answers. Importantly, similarity-based selection requires sufficient human response data to estimate cultural similarities, which may be difficult to obtain for low-resource cultures. In contrast, established cultural-sphere priors provide a practical alternative when such data are scarce.

In the revision, we will incorporate these results to further clarify the contribution and applicability of cultural-sphere priors.

W2: VSM Calibration and Evaluation Validity

We conducted an additional analysis to address this issue and further clarify the robustness of our findings.

(1) Offset-Adjusted Similarity. To address the calibration concern, for each model, VSM dimension, and target country, we estimated an additive constant C from the mean difference between published Hofstede and Vanilla scores across the other 36 countries, applying the same C to both methods. Under this leave-one-country-out calibration, CSP achieves an average similarity of 79.91, compared with 79.51 for Vanilla (+0.40 pp)

ModelVanillaCSP
DeepSeek-V3.278.5778.11 (−0.46)
GPT-4o79.2279.72 (+0.50)
GPT-5.279.1380.25 (+1.11)
Kimi-K2.580.2582.09 (+1.84)
Llama-3.3-70B79.0578.30 (−0.75)
Mistral-Large-380.3779.58 (−0.80)
Mistral-Medium-250579.4379.87 (+0.44)
o380.0481.34 (+1.30)
Mean79.5179.91 (+0.40)

(2) Existing Structural Evidence. Table 3 evaluates the consistency of pairwise cultural similarity rankings between models and human references. This metric is invariant to uniform additive offsets within each VSM dimension. CSP improves Spearman's correlation from 0.236 to 0.381, providing complementary evidence that does not depend on such calibration constants.

Together, these results further support the robustness of our findings. We will incorporate the additional analysis and clarify the limitations of absolute VSM score comparisons in the revision.

W3: Potential Judge Bias in PRISM Evaluation

We conducted an additional Blind Judge Evaluation to eliminate the potential influence of generation prompts. Specifically, we retained the original Vanilla and CSP responses, GPT-4o judge, PRISM references, and scoring rubric, while removing the generation system prompts from the judge's input.

ModelVanillaCSP
GPT-4o78.3479.54 (+1.5%)
GPT-4.178.5181.02 (+3.2%)
GPT-579.0680.87 (+2.3%)
GPT-5.277.9677.99 (+0.0%)
o376.5678.37 (+2.4%)
Kimi-K2.579.3379.47 (+0.2%)
DeepSeek-V3.277.8179.14 (+1.7%)
Mistral-Large-381.0881.33 (+0.3%)
Mistral-Medium-250579.6279.91 (+0.4%)
Llama-3.3-70B74.2175.40 (+1.6%)
Average78.2579.30 (+1.4%)
95% CI of relative gain—[+0.6%, +2.1%]

The overall improvement trend remains consistent: CSP achieves gains across all 10 models, with an average relative improvement of +1.4% (95% CI: [+0.6%, +2.1%]). This further supports the effectiveness of cultural-sphere prompting under controlled evaluation conditions. We will include these results in the revision.

W4: Impact of Proxy Supervision on Country-Specific Preferences

We address this concern by clarifying the role of consensus supervision and providing further empirical evidence.

(1) Effectiveness of Proxy Supervision for Cold-Start Alignment. When target-culture training data are unavailable, cultural-sphere consensus provides a practical source of proxy supervision. As shown in Table 5, SFT-Aug improves held-out WVS accuracy across all three Qwen3 model scales, demonstrating its effectiveness in this setting.

(2) Improvements Across Consensus Levels. Following your suggestion, we conducted additional experiments to compare the results across high- and low-consensus questions. As shown in the table below, SFT-Aug achieves substantial gains for held-out low-resource cultures on high-consensus questions (+12.08, +4.35, and +7.23 pp for 4B/8B/14B). Importantly, performance does not decline on low-consensus questions, with further gains of +0.61, +3.03, and +3.89 pp, respectively. This suggests that consensus supervision may facilitate the transfer of shared cultural patterns while maintaining reasonable performance on less-consensual questions.

ModelMethodHigh consensusLow consensus
Qwen3-4BVanilla35.9132.03
SFT-Naive43.44 (+7.53)33.28 (+1.25)
SFT-CSP45.84 (+9.93)30.19 (−1.85)
SFT-Aug47.99 (+12.08)32.64 (+0.61)
Qwen3-8BVanilla56.6832.60
SFT-Naive56.92 (+0.24)32.72 (+0.12)
SFT-CSP60.63 (+3.95)34.66 (+2.06)
SFT-Aug61.02 (+4.35)35.63 (+3.03)
Qwen3-14BVanilla66.1834.69
SFT-Naive66.63 (+0.46)34.12 (−0.57)
SFT-CSP68.54 (+2.36)33.22 (−1.47)
SFT-Aug73.40 (+7.23)38.58 (+3.89)
Reviewer_x1vhOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether relationships among cultures can improve cultural alignment in language models. It uses eight predefined cultural spheres in two ways: prompting models with answers from countries in the same sphere and contrasts from other spheres, and fine-tuning for countries without target-specific training labels using the consensus answer of other sphere members. The experiments cover ten models and three benchmarks: WVS, VSM, and PRISM.

Critical strengths

The paper turns a clear intuition about related cultures into concrete prompting and cold-start fine-tuning methods. The prompting evaluation is broad: SC-ICL improves WVS accuracy for all ten models. The ablations are also informative, showing that answers shared by same-sphere countries account for most of the prompting gain. The paper provides detailed prompts and data splits, and reports the unfavorable open-ended fine-tuning results in the appendix.

Critical weaknesses

The contribution of the predefined sphere structure is not isolated from that of the additional answer evidence. SC-ICL supplies the model with other countries’ answers as well as sphere information. With three demonstrations, adding same-sphere agreement raises WVS accuracy from 53.54% to 58.27%; adding the cross-sphere contrast raises it only to 58.51% (Figure 4). This establishes the value of the selected answer evidence, but does not show that the predefined spheres select better references than a more general similarity rule. A useful control would keep the number of reference answers and prompt length fixed while removing sphere labels. Another would select reference countries by answer similarity measured only on separate demonstration questions.

Consensus supervision may suppress answers distinctive to a held-out country. The fine-tuning method uses the modal answer of other sphere members as a pseudo-label for that country. The aggregate results in Table 5 show benefits in some settings, but do not reveal what happens when the held-out country’s human modal answer differs from the consensus of its sphere peers. The existing high- and low-consensus analysis concerns prompting and does not directly resolve this question for fine-tuning. I would like to see held-out test results split by whether the target country agrees with its sphere peers, before and after fine-tuning. This would show whether the method transfers shared patterns while preserving genuine country-level differences.

Questions for authors

How much of the prompting gain comes from sphere labels versus the selected answers? What happens when references are selected using similarity estimated from separate demonstration questions?

1 comment
AuthorsOpenReview ↗

Response to Reviewer

Thanks for your supportive review and valuable suggestions. Our responses to the weaknesses (W) and questions (Q) are provided below. A revision of our paper will be uploaded, with the revised parts marked in blue.

W1, Q1, Q2: Contributions of Cultural-Sphere Priors vs. Selected Answers

We would like to clarify the contribution of cultural-sphere priors versus the selected answers through existing and additional experimental evidence.

(1) Existing Evidence. Table 1 shows that adding cultural-sphere context without any reference answers (Vanilla → CSP) yields average relative gains of +1.5%, +3.7%, and +3.3% on WVS, VSM, and PRISM, respectively. On WVS, with the same demonstration questions and target-country answers, adding sphere context raises accuracy from 52.89% to 53.54% (Standard ICL → SP-ICL), while further incorporating sphere-based reference answers raises it to 58.51% (SC-ICL).

(2) Additional Reference Selection Ablation. To examine whether predefined spheres provide useful guidance beyond simply adding reference answers, we compared three strategies on GPT-4o, GPT-5, and GPT-5.2, keeping the demonstration questions, target-country demonstration answers, and number of references fixed.

  • Random selection samples the same number of reference countries as SC-ICL from countries with answers to each demonstration question, without using sphere membership.
  • Similarity-based selection ranks countries by the similarity of their human response distributions on separate training questions, then selects the closest countries with answers to each demonstration question, without accessing test answers.
  • Sphere-based corresponds to SC-ICL.
ModelSphere-basedRandomSimilarity-based
GPT-4o60.66%57.75%58.62%
GPT-559.52%56.72%55.41%
GPT-5.257.46%55.05%55.05%
Average59.21%56.51%56.36%

Sphere-based selection outperforms both alternatives across all three models, supporting the value of cultural relationships beyond additional reference answers. Importantly, similarity-based selection requires sufficient human response data to estimate cultural similarities, which may be difficult to obtain for low-resource cultures. In contrast, established cultural-sphere priors provide a practical alternative when such data are scarce.

In the revision, we will incorporate these results to further clarify the contribution and applicability of cultural-sphere priors.

W2: Does Consensus Supervision Suppress Country-Specific Preferences?

We address this concern by clarifying the role of consensus supervision and providing further empirical evidence.

(1) Effectiveness of Proxy Supervision for Cold-Start Alignment. When target-culture training data are unavailable, cultural-sphere consensus provides a practical source of proxy supervision. As shown in Table 5, SFT-Aug improves held-out WVS accuracy across all three Qwen3 model scales, demonstrating its effectiveness in this setting.

(2) Improvements Across Consensus Levels. Following your suggestion, we conducted additional experiments to compare the results across high- and low-consensus questions. As shown in the table below, SFT-Aug achieves substantial gains for held-out low-resource cultures on high-consensus questions (+12.08, +4.35, and +7.23 pp for 4B/8B/14B). Importantly, performance does not decline on low-consensus questions, with further gains of +0.61, +3.03, and +3.89 pp, respectively. This suggests that consensus supervision may facilitate the transfer of shared cultural patterns while maintaining reasonable performance on less-consensual questions.

ModelMethodHigh consensusLow consensus
Qwen3-4BVanilla35.9132.03
SFT-Naive43.44 (+7.53)33.28 (+1.25)
SFT-CSP45.84 (+9.93)30.19 (−1.85)
SFT-Aug47.99 (+12.08)32.64 (+0.61)
Qwen3-8BVanilla56.6832.60
SFT-Naive56.92 (+0.24)32.72 (+0.12)
SFT-CSP60.63 (+3.95)34.66 (+2.06)
SFT-Aug61.02 (+4.35)35.63 (+3.03)
Qwen3-14BVanilla66.1834.69
SFT-Naive66.63 (+0.46)34.12 (−0.57)
SFT-CSP68.54 (+2.36)33.22 (−1.47)
SFT-Aug73.40 (+7.23)38.58 (+3.89)
Reviewer_SyD2OpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors investigate whether cultural-sphere structure can serve as a relational prior for cultural alignment. They show that across 10 models and 3 complementary benchmarks, sphere-aware prompting consistently outperforms culture-only baselines. They ablate the prompt and find agreement among same-sphere cultures as the strongest effective signal. They show that using within-sphere consensus as proxy supervision, sphere-aware fine-tuning substantially improves alignment on held-out cultures and generalizes to an unseen benchmark.

Critical strengths
  • The relational view of cultural alignment is interesting, and the cold-start setting is relevant for low-resource cultures.
  • The evaluation is extensive covering 10 models and 3 benchmarks with different formats.
  • The authors validate that the spheres reflect real structure in human data.
  • The appendix is detailed, and negative results are reported.
Critical weaknesses
  • An ablation where random cultures are added to the prompt is missing. SC-ICL adds survey answers from other countries, so it is unclear if the gain comes from the sphere structure or just from the extra data.
  • There is no simple non-LLM baseline. Based on Table 12, copying the answers of a random same-sphere country already gives higher WVS accuracy than any prompting method in Table 1.
  • The spheres are derived from WVS data, and the paper then evaluates on WVS. This circularity should be discussed.
  • What is the variance in each result? Is the gain significant or within stochastic error? The CIs in Table 1 are computed over the ten models, but each model is run only once, and the decoding settings are not reported. I am worried the results might not be significant if generation's stochasticity is accounted for.
  • The language should be tuned down. For example, CSP gives only +1.5% on WVS, its CI crosses zero, and it hurts some models. So “consistently outperforms” is overstated.
  • The finding that same-sphere agreement helps most on high-consensus questions is expected by definition. On these questions all sphere members share the same answer, so copying them is always correct.
  • The low-resource cultures start from a lower baseline, so larger gains are expected.
  • The culture spheres group countries that are culturally very different. For example, African-Islamic contains Nigeria, Iran, and Indonesia, and West & South Asia contains India, the Philippines, and Vietnam. The paper’s own human data places Morocco and the Philippines closer to other spheres. This makes me question the premise of the paper.
  • Figure 3 ticks and legend are impossible to read.
Questions for authors

Please address weakness above

7 comments
AuthorsOpenReview ↗

Response to Reviewer

Thanks for your supportive review and valuable suggestions. Our responses to each Weaknesses (W) and Questions (Q) are as follows. A revision of our paper will be uploaded, with the revised parts marked in blue.

W1: Missing Random References Ablation

Following your suggestion, we conducted an additional Random References Ablation on three representative models (GPT-4o, GPT-5, and GPT-5.2) to isolate the contribution of cultural-sphere-based reference selection.

Experimental Setup. We kept the demonstration questions, target-country demonstration answers, and number of references unchanged, while replacing sphere-based reference selection with random sampling from other countries.

Results. Random References consistently underperforms SC-ICL across all three models. These results suggest that the benefits of SC-ICL extend beyond simply providing additional cross-cultural answers, supporting the value of cultural sphere-informed reference selection.

ModelSC-ICLRandom ReferencesDifference (pp)
GPT-4o60.66%57.75%−2.91
GPT-559.52%56.72%−2.80
GPT-5.257.46%55.05%−2.41
Average59.21%56.51%−2.71

W2: Missing Non-LLM Baseline

We would like to clarify this issue from two aspects.

(1) The proposed copying baseline requires information unavailable in our setting and would not constitute a fair comparison. It requires human answers from a same-sphere country to each test question being evaluated. In our setting, however, human answers are available to the methods only for 50 demonstration questions, strictly disjoint from the 149 evaluation questions (Appendix E.1). SC-ICL uses cross-cultural answers to these demonstration questions to predict answers to unseen evaluation questions. Direct copying would therefore not provide a comparison under the same information constraints. Its applicability is also limited to questions for which another country’s human answers have already been collected; it does not address generalization to new questions without such answers. Our method instead investigates whether cultural-sphere information can support this generalization, which is more useful in practice.

(2) Table 12 aims to provide evidence of shared cultural patterns, motivating our approach. It measures agreement between countries’ human answers to examine whether cultural spheres capture shared value patterns. As you observed, the high within-sphere agreement indicates meaningful similarities among related cultures. This observation motivates our investigation of cultural-sphere structure as a relational prior.

We will clarify this distinction in the revision.

W3: Potential Circularity Between Cultural Spheres and WVS Evaluation

We would like to clarify this concern from three aspects below.

(1) Established Prior, Not Evaluation-Driven Spheres. The Inglehart–Welzel cultural spheres that we used through the study are a well-established sociological framework developed with references from successive WVS waves and broader historical and cultural considerations. However, they are not directly constructed or optimized using our WVS evaluation questions.

(2) Cross-Dataset Validation. We further validate this prior using both WVS and Hofstede 6-D (Appendix D). All eight spheres show higher within-sphere similarity (average ratios: 1.172 and 1.167), with clustering agreement significantly above chance (ARI: 0.245 and 0.420; p = 0.0001). This supports the validity of the cultural-sphere prior across different cultural datasets.

(3) Research Scope and Limitations. Our goal is to examine whether this prior improves LLM cultural alignment, not to re-establish the underlying theory. Improvements on VSM (+3.7%) and PRISM (+3.3%) have provided evidence for its effectiveness beyond WVS. We also explicitly acknowledge the reliance on a single cultural taxonomy and the fixed 2023 cultural map in Appendix A.

W4: Statistical Significance and Reproducibility

Thank you for raising this concern and for recognizing our existing statistical analysis. Table 1 reports 95% confidence intervals across ten models, while Appendix H.1 provides additional culture-level bootstrap confidence intervals, further supporting the reliability of our findings. We provide the following additional evidence and clarifications.

(1) Additional Repeated-Run Evaluation. We selected three representative models (GPT-4o, GPT-5, and GPT-5.2) to further assess decoding variability. For each model, we repeated both Standard ICL and SC-ICL three times on 149 WVS questions across 41 countries. We computed 95% confidence intervals using 30,000 country-level bootstrap resamples and assessed significance using 200,000 two-sided paired sign-flip permutations at the country level. SC-ICL consistently outperforms Standard ICL across all three models.

ModelStandard ICLSC-ICLRelative Gain (95% CI)
GPT-4o54.38%60.27%+10.83% [8.99%, 12.84%]
GPT-553.69%59.05%+9.98% [8.19%, 11.69%]
GPT-5.253.47%57.04%+6.67% [5.36%, 8.09%]

(2) Decoding Settings. We used a temperature of 1.0, with model-specific output limits of 128 tokens or 5,128 completion tokens. We will explicitly report these settings in the revision.

Together, these additional analyses and clarifications further support the statistical reliability and reproducibility of our findings.

W5: Overstatement of “Consistently Outperforms”

Thank you for this constructive suggestion, which helps improve the precision of our presentation.

Broad Improvements Across Models and Benchmarks. Our results demonstrate broad improvements rather than isolated gains. As shown in Table 1, CSP improves alignment on 8/10, 9/10, and 10/10 models across WVS, VSM, and PRISM, respectively. Furthermore, SC-ICL improves WVS accuracy across all 10 models, with an average relative gain of +12.1%.

Wording Refinement. To better reflect these findings, we will replace “consistently outperforms” with “generally outperforms” in the general claim, while retaining “consistently” specifically for SC-ICL on WVS.

W6: Is the High-Consensus Finding Trivial?

We would like to clarify that the stronger gains on high-consensus questions support the motivation of our cultural-sphere approach but they do not represent a trivial consequence of the consensus definition.

(1) No Access to Cross-Culture Consensus on Test Questions. SC-ICL receives cross-cultural answers only for 50 separate demonstration questions, strictly disjoint from the 149 evaluation questions (Appendix E.1). The High-/Low-Consensus grouping is computed from human ground-truth for analysis, while neither the grouping nor the test-question consensus answers are supplied to the model or used to select demonstrations. Thus, the model cannot directly copy a same-sphere answer to the test question. Any benefit must transfer across different questions, rather than follow automatically from the consensus definition.

(2) Validation of Our Motivation. Our method is motivated by the idea that shared cultural patterns within a sphere can provide useful guidance for cultural alignment. The larger improvements on high-consensus questions (Figure 4) provide empirical support for this hypothesis, as the method benefits most where such shared patterns are stronger.

In the revision, we will clarify that the demonstration and evaluation questions are strictly disjoint, and that the High-/Low-Consensus grouping is used solely for analysis. We will emphasize that the observed gains on unseen questions suggest generalization of shared cultural patterns from separate demonstrations, rather than direct copying of test-question consensus answers.

W7: Are Larger Low-Resource Gains Simply Due to Lower Baseline Performance?

We agree that lower baseline accuracy provides more room for improvement, so larger absolute gains alone are insufficient to establish a greater benefit for low-resource cultures. We therefore consider two complementary comparisons.

(1) Lower baseline accuracy alone does not guarantee large gains. On the same low-resource cultures, starting from the same Vanilla accuracy of 50.15%, Standard ICL improves accuracy by only 0.76 pp, whereas SC-ICL improves it by 7.52 pp (Table 27). Thus, the available improvement headroom alone does not explain the magnitude of the SC-ICL gains, highlighting the additional benefits of leveraging cultural-sphere relationships beyond conventional demonstrations.

(2) The advantage remains after headroom normalization. We compute relative error reduction as the accuracy gain divided by the Vanilla error rate. SC-ICL reduces errors by 15.1% for low-resource cultures, compared with 12.3% for high-resource cultures, using the accuracies in Table 4. Thus, the larger low-resource gains persist under this normalization accounting for differences in initial accuracy.

We will add these explanations in the revision.

W8: Cultural Heterogeneity Within Spheres

We agree that countries within a cultural sphere can differ substantially. Our premise is that these spheres capture useful shared patterns, not that their members are culturally homogeneous or that every country must be closest to its assigned sphere.

(1) Well-Established Cultural Prior Rather Than Homogeneous Groups. We adopt the Inglehart–Welzel framework, developed through decades of sociological research, rather than constructing a new classification. Such broad groupings can capture shared value patterns while retaining substantial country-level differences and cross-sphere overlap. Our focus is on whether these coarse-grained relationships capture useful cultural similarities for LLM alignment, without requiring that countries within a sphere must share identical values.

(2) Empirical Evidence of Both Shared Structure and Practical Utility. Appendix D (Table 12) shows that all eight spheres exhibit higher within-sphere than between-sphere similarity on both WVS and Hofstede 6-D. Even for African-Islamic and West & South Asia, the WVS ratios reach 1.153 and 1.096, respectively. These results support the presence of meaningful shared patterns despite internal differences. Importantly, our alignment experiments also validate this coarse structure is useful. These findings support the practical value of this prior.

(3) Heterogeneity and Limitations. We do not overlook country-level differences. Appendix D (Figures 10–11) illustrates discrepancies between predefined spheres and data-driven clusters, while Appendix A explicitly discusses the limitations of relying on a single cultural taxonomy.

Therefore, our contribution lies in demonstrating that even coarse-grained cultural spheres can capture useful shared patterns for alignment, rather than assuming they are culturally homogeneous. We will enhance the clarification in the revision.

W9: Readability of Figure 3

Thank you for this suggestion. Figure 3 is intended to illustrate the overall cultural similarity patterns, rather than emphasize individual values in the color scale. Larger and clearer visualizations of the related cultural structures are provided in Appendix D (Figures 10–11) for reference.

Nevertheless, we will improve the figure size, font readability, and legend layout in the revision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

The revision of our work has been uploaded. Looking forward to your futher feedback!

Dear Reviewers,

Thank you for your thoughtful and constructive feedback. Following your suggestions, we have revised the paper and uploaded the updated version, with revisions marked in blue. The revision incorporates the additional experimental results, methodological clarifications, and discussions described in our responses.

We hope these changes address your concerns and make our work clearer. We would greatly appreciate any further comments or questions you may have during the discussion period.

Thank you again for your time and consideration.

Best regards,
The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.