PaperlayerView exact revision on arXiv

Source-linked AI summary

CoCoA: Context-Conditional Cultural Alignment for Large Language Models

Kyungdon Lee, Wei Xu, Alan Ritter, Dong-Ho Lee, JinYeong Bak

arXiv:2608.29492v1cs.CL

TL;DR

LLMs often favor Western-associated entities, while conventional uniform debiasing can remove culturally appropriate associations. CoCoA uses dual-context training and complementary objectives to condition preferences on cultural cues, reducing measured bias while preserving neutrality in uncued contexts.

  • Problem

    Existing mitigation methods often target uniform neutrality, leaving open how to reduce inappropriate cultural preferences without removing associations that are appropriate in context.

  • Method

    CoCoA learns context-conditional preferences through dual-context training on the same entity pairs, complementary loss functions, and goal-aware gradient optimization.

  • Results

    CoCoA reduces the Cultural Bias Score from 43 to 24 on average across four models and ten languages while maintaining near-neutral preferences at 50.2.

  • Takeaways & Limitations

    Context-sensitive mitigation better distinguishes when culturally supported associations are appropriate from when balanced behavior is desired.

  • Takeaways & Limitations

    The evaluation uses binary cultural cues and entity-centric fill-in-the-blank tasks, not implicit or graded cues, open-ended generation, or downstream cultural reasoning.

Abstract

from arXiv · show

Large Language Models (LLMs) often favor Western-associated entities across cultural contexts. Conventional debiasing methods aim for uniform neutrality, but cultural bias mitigation demands context-conditional behavior, preferring culturally appropriate entities when cultural cues are present and remaining neutral when they are absent. We propose CoCoA (Context-Conditional Cultural Alignment), a framework that learns this behavior through dual-context training on the same entity pairs under contexts with and without cultural cues. CoCoA combines a contrastive alignment objective with calibration and drift regularization, optimized through goal-aware gradient reconciliation. We evaluate CoCoA on CAMeL and Camellia, two entity-centric cultural bias benchmarks, across ten language settings and four LLMs. CoCoA reduces the Cultural Bias Score from 43 to 24 on average while maintaining near-neutral preferences at 50.2, with minimal impact on general performance across five standard benchmarks. These findings highlight that effective cultural alignment requires context-conditional modeling rather than uniform debiasing, and establish a new direction for mitigating entity-centric cultural bias in LLMs.

1 Introduction

LLMs often favor Western-associated entities, while uniform debiasing can erase culturally appropriate associations. CoCoA addresses this mismatch by learning culturally appropriate preferences when cues are present and neutrality when they are absent.

  • LLMs often default to Western cultural preferences across languages and cultural settings, including entity-level choices.
  • Uniform debiasing can flatten culturally appropriate associations together with inappropriate ones, producing cultural homogenization.
  • CoCoA trains on paired cued and uncued contexts using the same culturally associated entity pair.
  • CoCoA encourages culturally associated entities under explicit cultural cues while balancing entity preferences without those cues.
  • 43 to 24: CoCoA reduces the Cultural Bias Score on average across four models and ten languages while maintaining near-neutral preferences at 50.2.
  • The evaluation covers seven cultural groups, ten language settings, and four model families, with consistent gains over adapted social bias baselines while maintaining general model performance.

2 Related Work

Prior research documents Western-centric tendencies and develops cultural-alignment and social-bias mitigation methods. CoCoA targets the mismatch between uniform neutrality and context-dependent cultural appropriateness.

  • Prior studies document Western-centric tendencies in cultural values, commonsense reasoning, cultural knowledge perception, and entity-level preferences.
  • Cultural-alignment methods include culture-specific training data, self-pluralising alignment, steering, preference optimization, grounded cultural knowledge, contrastive learning, and expert routing.
  • Social-bias mitigation spans approaches such as knowledge editing that equalizes stereotypical and antistereotypical associations.
  • Uniform-neutrality objectives are misaligned with cultural settings because they can erase legitimate cultural associations, motivating separate objectives for cued and uncued contexts.

3 CoCoA

CoCoA formulates cultural bias mitigation as context-conditional entity preference: favor culturally associated entities when cues are present while keeping preferences balanced otherwise. It learns this behavior by scoring the same contrasting entity pair in grounded and neutral contexts, combining complementary losses, regularization, and gradient reconciliation.

  • Problem formulation: CoCoA targets local-entity preference in culturally grounded contexts and balanced preferences in culturally neutral contexts.
  • Entity scoring: Prior-adjusted scores subtract an entity’s vanilla-model likelihood to isolate contextual contributions to entity preference.The correction strength α is configurable and is fixed at different training values for grounded and neutral contexts.
  • Dual-context training: Each training instance uses the same local–other-culture entity pair under one grounded and one neutral context.The local entity is associated with the cued culture, while the counterpart comes from another culture.
  • Complementary objectives: The grounded contrastive loss encourages the local entity to outrank its counterpart, while neutral calibration encourages similar scores.The pairwise grounded objective uses the prior-adjusted score difference, with τ controlling preference sharpness.
  • Neutral stability: Drift regularization penalizes deviations of the neutral score gap from its vanilla value during joint optimization.Together with neutral calibration, it stabilizes neutral-context behavior while the grounded objective updates shared LoRA parameters.
  • Optimization: Goal-aware PCGrad reconciles weighted grounded and neutral-side gradient directions before updating the LoRA adapters.When directions conflict, projection strength depends on the opposing objective’s distance from its target; otherwise gradients remain unchanged.

4 Experiments

CoCoA is evaluated on two entity-centric cultural bias benchmarks across ten language settings, four backbone models, and five-fold cross-validation. It lowers grounded-context bias while generally preserving neutral calibration, whereas neutralization baselines fail to reliably separate the two contexts.

  • Datasets and setup: CAMeL and Camellia evaluate culturally associated entity pairs across grounded and neutral contexts, covering seven cultural groups and ten language settings.CAMeL focuses on Arabic, while Camellia covers Asian language settings; grounded labels indicate unique target-culture appropriateness.
  • Datasets and setup: The study evaluates Llama-3.1-8B, Qwen3-8B, Mistral-7B-v0.3, and Gemma-3-12B-pt.Qwen3-8B is included to assess generalization beyond pretrained backbones to an instruction-tuned model.
  • Datasets and setup: Five-fold cross-validation separates contexts, entities, and entity pairs across training, validation, and evaluation splits.Evaluation uses entities the model has not been trained on.
  • Main results: CBSg falls from 43 to 24 on average across ten language settings and four backbones, while CBSn reaches 50.2.Lower CBSg indicates stronger local preference in grounded contexts, while CBSn near 50 indicates neutral balance.
  • Main results: CoCoA consistently reduces CBSg across all four backbone families while keeping CBSn near 50 for Llama-3.1-8B, Qwen3-8B, and Mistral-7B-v0.3.Gemma-3-12B-pt shows a larger neutral-calibration gap, with CBSn at 45.6.
  • Main results: Neutralization baselines do not produce context-conditional behavior, often achieving separation by moving neutral-context behavior away from 50.For Llama-3.1-8B, BiasUnlearn changes CBSg from 33.1 to 37.6 and CBSn reaches 44.7; BiasEdit shifts Korean CBSn from 49.4 to 21.8 while lowering CBSg from 47.0 to 20.8.

5 Analysis

CoCoA shifts entity preferences toward local entities in culturally grounded contexts while preserving balanced behavior in neutral contexts, with only marginal changes on general benchmarks. The analysis finds improvements across entity types rather than a narrow subset.

  • Objective ablation and weight sensitivity: CBSg rises to 54.6 when the grounded loss is removed, confirming that Lg controls local preference in grounded contexts.
  • Per-pair preference shift: 79% of grounded-context pairs shift toward stronger local preference under CoCoA, as shown by points above the diagonal.
  • Per-entity-type breakdown: Negative ∆CBSg values occur across all language–entity type combinations, with the largest reductions for Korean, Japanese, and Arabic.
  • Per-entity-type breakdown: Authors and Names show particularly large preference shifts within some language settings.
  • General modeling performance: The five-benchmark evaluation averages CoCoA performance across 50 checkpoints spanning ten language settings and five folds.
  • General modeling performance: CoCoA changes general benchmark performance only marginally; the largest drop is 0.38 percentage points on ARC-Challenge, and remaining changes are below 0.32 points.

6 Conclusion and Future Work

The paper argues that cultural bias mitigation should distinguish contexts where cultural associations are appropriate from contexts requiring balance. CoCoA implements this distinction through dual-context training and related objectives, with evaluations showing reduced grounded-context Western preference while preserving neutral-context balance.

  • Conclusion: CoCoA induces context-conditional cultural preferences through dual-context training, complementary loss functions, and goal-aware gradient optimization.
  • Conclusion: Across seven cultural groups, ten language settings, and four backbone models, CoCoA reduces Western preference in grounded contexts while preserving balance in neutral contexts.
  • Future work: The current binary local–other-culture pairing does not address multicultural alignment among multiple non-Western cultures.

Limitations

CoCoA’s performance and practical deployment are constrained by configuration-dependent weight sensitivity and uneven improvement across settings. Its evaluation also covers only explicit binary cues, local–Western pairings, and entity-centric completion tasks.

  • Configuration sensitivity: CoCoA’s weight setting does not transfer uniformly across configurations, with seven of forty model–setting configurations using non-default weights.
  • Evaluation scope: The formulation treats cultural-cue presence as binary and does not handle implicit or graded cues.
  • Evaluation scope: Both benchmarks pair local entities against Western counterparts and do not establish behavior when multiple non-Western cultures are aligned simultaneously.
  • Evaluation scope: The evaluation focuses on entity-centric fill-in-the-blank tasks and does not cover open-ended generation or downstream cultural reasoning.

Ethics Statement

The ethics statement frames CoCoA as a context-sensitive fairness method intended to preserve legitimate cultural associations and balanced preferences. It also cautions that the method could be misused to steer models toward specific cultural viewpoints.

  • Ethical aims: CoCoA aims to make entity preferences more sensitive to context while preserving culturally appropriate associations and balanced preferences in neutral contexts.
  • Ethical aims: The method is not intended to erase genuine cultural associations or suppress legitimate cultural expressions.
  • Responsible use: The authors encourage responsible use because bias mitigation methods can be misused to steer models toward specific cultural viewpoints.

A.1 Implementation Details

CoCoA uses context-conditional training and calibrated objectives to distinguish culturally appropriate preferences from neutral-context preferences. Evaluations compare it with adapted social-bias baselines and prompting methods across languages and models.

  • Objective Weights: Default objective weights are wg=1.0, wn=2.0, and wd=1.0; most configurations retain them.Thirty-three of 40 model–setting configurations use the default, while seven receive validation-based adjustments.
  • Objective Weights: Adjusted wd moves neutral-context CBSn toward 50 when default weights produce directional drift.The adjustment raises wd when CBSn exceeds 50 and lowers it when CBSn falls below 50.
  • Baselines: CoCoA adapts BiasEdit and BiasUnlearn by replacing social-group pairs with local–other-culture entity pairs and training on neutral contexts.BiasEdit edits MLP representations, whereas BiasUnlearn uses forget and retain objectives.
  • Culture-Aware Prompting Baselines: On Llama-3.1-8B, CSP and CCT leave grounded CBSg near vanilla values, while CoCoA reaches 15.2 and 24.9 on two backbones.For Llama-3.1-8B, CSP is 36.2 and CCT 36.9 versus vanilla 36.4.
  • Culture-Aware Prompting Baselines: Prompting also lowers neutral-context CBSn, while CoCoA reaches 50.3 and 50.7; a single prompt cannot separate the two objectives.On Llama-3.1-8B, CSP and CCT reduce CBSn from 43.4 to 40.5 and 42.1.
  • Culture-Aware Prompting Baselines: Prompting baselines condition every query with cultural instructions, whereas CoCoA trains context-conditional behavior and adds no inference-time prefix.CSP and CCT use 24- and 39-word prefixes, respectively.
  • Joint Multicultural Adapter: A pooled adapter reduces grounded CBSg from 36.4 to 17.2 but overshoots neutral CBSn to 56.2 versus 50.3 for setting-specific adapters.The remaining neutral-calibration gap is consistent with interference between settings sharing one adapter, though calibration and pooling effects are not separated further.

B.4 Effect on Social Bias Metrics

CoCoA’s evaluation examines whether cultural alignment affects measured social bias and whether fine-tuning changes the priors used for PMI-based scoring. The reported results show no increase in social bias and only small CBS changes under re-estimated priors.

  • Social Bias Metrics: CoCoA training produces no increase in measured social bias on CrowS-Pairs or StereoSet.The evaluation uses CrowS-Pairs stereotypical preference rate and StereoSet Stereotype Score, LMS, and ICAT.
  • Social Bias Metrics: CrowS-Pairs decreases from 65.8 to 65.3, while StereoSet SS changes from 66.40 to 66.23 after CoCoA training.LMS and ICAT change only marginally.
  • Prior Drift: CoCoA re-estimates entity priors with the vanilla BOS-only protocol and compares CBS using vanilla versus fine-tuned priors.The analysis averages results across ten language settings and five folds.
  • Prior Drift: Llama-3.1-8B has mean absolute prior drift of 0.45 log-probability units, while Qwen3-8B has 1.76 units.Qwen3-8B therefore exhibits approximately 3.9× larger drift.
  • Prior Drift: Switching to fine-tuned priors changes average CBSg by 1.1 points for Llama and 2.4 points for Qwen.Llama changes from 15.2 to 14.1, while Qwen changes from 24.9 to 22.5.
  • Interpretation: CoCoA’s LoRA adapters alter attention patterns for context-conditional preference without substantially changing unconditional entity priors.This supports reusing priors computed from the vanilla model for PMI evaluation.

D.2 Gradient Combination

The gradient projection is tested as a safeguard against interference rather than as the source of CoCoA’s improvement. Removing it leaves grounded and neutral CBS results essentially unchanged.

  • Gradient Combination: Projection and plain weighted-sum training differ by only 0.1 in CBSg and 0.4 in CBSn, both within reported standard errors.The two runs use identical losses, weights, data, splits, epochs, and optimizer settings.
  • Gradient Combination: Both gradient-combination methods reduce CBSg from 36.4 to roughly 15 and bring CBSn from 43.4 to within half a point of 50.The reported comparison attributes the context-conditional behavior to the dual-context objective rather than gradient surgery.
  • Interpretation: The projection is retained because it is inexpensive and guards against interference in configurations not swept, despite showing no advantage in the reported comparison.The per-setting weights were selected in projection-based runs, limiting the comparison’s symmetry.

E.1 Score Decomposition

Score decomposition indicates that CoCoA’s grounded-context gains mainly come from lowering inappropriate other-culture scores rather than broadly promoting local entities. Entity-level results show the context-sensitive pattern across most held-out entities, with no formal guarantee for every pair.

  • Grounded-Context Decomposition: In grounded contexts, other-culture scores fall by 6.87 while local scores rise by 0.21.Local scores rise in 54.8% of grounded observations, close to chance, so the widened gap mainly reflects ruling out inappropriate alternatives.
  • Neutral-Context Decomposition: In neutral contexts, other-culture scores increase by 1.26 versus 0.38 for local scores, accompanying CBSn’s movement from 43.4 to 50.3.This is consistent with correcting the vanilla model’s under-ranking of other-culture entities without a cultural cue.
  • Interpretation: The combined observations indicate cue-conditioned re-ranking rather than general aversion to Western entities.The effect is limited to contexts where the explicit cue makes the other-culture entity inappropriate, while neutral contexts move toward balance.
  • Entity-Level Context Sensitivity: Across counterparts, Δg > Δn holds for 88.5% of an entity’s comparisons, and 83.1% of entities satisfy this for at least four of five counterparts.The mean gap shift is +6.95 with a cue and −0.60 without one.
  • Entity-Level Context Sensitivity: The entity-level proportions are computed over held-out entities and counterparts, without a formal guarantee that the ordering holds for every pair.The result supports broad consistency but remains an empirical proportion rather than a universal claim.
Loading 2608.29492v1…