Same Values, Different Languages? From Multilingual Probing to Steering LLMs Toward Chinese Social Values

Yuemei Xu; Kexin Xu; Jian Zhou; Haoyu Lu; Yequan Wang; Aishan Liu

Source-linked AI summary

Same Values, Different Languages? From Multilingual Probing to Steering LLMs Toward Chinese Social Values

Yuemei Xu, Kexin Xu, Jian Zhou, Haoyu Lu, Yequan Wang, Aishan Liu

arXiv:2609.08515v1Sep 8, 2026cs.CLcs.AI

TL;DR

The paper examines whether LLMs maintain consistent preferences for culturally grounded Chinese Social Values across languages. It constructs multilingual contrastive probes and uses hidden-state-derived value vectors for fine-tuning-free, selective inference-time steering. Across six languages, preferences vary by model and language, while steering is effective, transfers cross-lingually, and generalizes to existing value benchmarks.

Problem
Whether LLMs exhibit consistent Chinese Social Value preferences across languages remains insufficiently examined, especially in realistic multilingual dilemmas.
Method
The paper constructs six-language C-Voices contrastive probes and derives value vectors from hidden-state discrepancies for selective inference-time steering.
Results
Experiments show model-dependent, language-sensitive CSV preferences and effective steering with cross-lingual transfer and generalization to existing value benchmarks.
Takeaways & Limitations
CSV-oriented behavior is not uniform across models and languages, but representation-level steering can improve CSV alignment while largely preserving general utility.

Abstract

from arXiv · show

As Large Language Models (LLMs) are increasingly integrated into human society, aligning them with pluralistic social values has become a critical priority. However, whether LLMs exhibit consistent value preferences across languages remains underexplored, particularly for culturally grounded values, which are more abstract and difficult to evaluate and align than safety-centric principles. We investigate this issue through Chinese Social Values (CSV), a value system rooted in Chinese culture and comprising $12$ dimensions across national, societal, and personal levels. We construct C-Voices, the first comprehensive multilingual contrastive probe dataset for CSV, with 86,400 dilemma-based instances in six languages, each pairing a CSV-aligned action with a value-conflicting alternative. Building on the contrastive probes of C-Voices, we then propose a fine-tuning-free value vector steering method that derives value directions from hidden-state discrepancies and selectively intervenes on value-sensitive layers during inference. Experiments on six languages show that CSV-oriented preferences are model-dependent and language-sensitive, with the same dilemma eliciting divergent responses across languages. Our method achieves effective CSV steering, supports cross-lingual transfer of value vectors, and generalizes to existing FLAMES and ValuePrism.

Introduction

The paper addresses the underexplored problem of whether LLMs express Chinese Social Values consistently across languages and develops a multilingual framework to evaluate and steer these behaviors.

Chinese Social Values comprise 12 culturally grounded dimensions organized across national, societal, and personal levels.
LLMs may respond differently to the same CSV-oriented dilemma across languages, raising questions about multilingual behavioral stability.
Existing steering approaches rely mainly on data-intensive reinforcement learning or supervised fine-tuning, while CSV lacks large-scale annotated data.
The proposed method derives value vectors from hidden-state discrepancies and selectively steers value-sensitive layers during inference.
C-Voices provides the first contrastive CSV probe dataset across six languages, covering all 12 dimensions with 86,400 dilemma-based instances.Each instance pairs contrasting behavioral choices to examine value-oriented decision-making.
Experiments show that CSV-oriented preferences are model-dependent and language-sensitive, while Chinese-derived value vectors transfer across five non-Chinese languages and generalize to existing value benchmarks.

The Proposed Framework

The framework constructs contrastive multilingual C-Voices to represent Chinese Social Values, then derives value directions and selectively injects them into value-sensitive layers during inference. Its representation-level design uses aligned-versus-conflicting behavior contrasts and layer selectivity to steer values while limiting unintended effects.

Framework Overview: The framework has two parts: constructing C-Voices to learn value-specific representations and identifying and selectively injecting value vectors into discriminative layers.The framework performs these operations at inference time rather than through fine-tuning.
C-Voices Dataset Construction: C-Voices grounds 12 Chinese Social Value dimensions across national, societal, and personal levels and builds dilemmas from contrasting value orientations.Conflict values are derived from Schwartz’s structured value space and supplemented with Chinese-context conflicts.
C-Voices Dataset Construction: Each C-Voices instance pairs a dilemma scenario with a value-aligned choice and a conflicting choice, enabling representation differences between aligned and conflicting behaviors to be measured.The aligned choice upholds the target value, whereas the conflicting choice prioritizes short-term or personal gain at its expense.
C-Voices Dataset Construction: The dataset pipeline selects conflict values, extracts topics from real-world contexts, and generates behavioral choices before manual quality filtering.The construction uses topic deduplication and retains 14,400 high-quality Chinese instances after review, with Cohen’s κ of 0.800 on a blind cross-check.
Value Representation Analysis: PCA shows separable hidden-state clusters for value-aligned and value-conflicting options, with separation varying across layers.This motivates steering through representation-level intervention focused on layers with stronger value-sensitive separability.
Value Steering Method: The method defines value directions from differences between aligned and conflicting representation means, selects the top-K sensitive layers, and injects directions conditionally during inference.Steering is gated by representation alignment with the value direction and activated only when the similarity exceeds a predefined threshold τ.

Experiment

Experiments across four LLMs and multiple languages show that CSV behavior and steering effectiveness depend on both model and language, while the proposed method transfers beyond C-Voices with limited utility degradation.

Setup: Experiments evaluate four LLMs using Support Rate and Likert Score across multilingual C-Voices.The models are Qwen3-8B, Qwen2.5-32B, LLaMA-8B, and Mistral-7B.
CSV-oriented behavior: CSV-oriented behavior is model-dependent, with distinct support distributions across the 12 dimensions.Qwen models show weaker support on Freedom than LLaMA-8B, while larger Qwen2.5-32B models outperform Qwen2.5-7B in supplementary comparisons.
CSV-oriented behavior: CSV-oriented behavior is also language-sensitive: LLaMA-8B is weaker in Arabic, whereas Mistral-7B reaches its lowest Likert Score in Chinese.Qwen models show more consistent behavior across the six languages than LLaMA-8B and Mistral-7B.
Steering results: +5.87 average Likert Score improvement is achieved by the proposed method, exceeding Causal intervention at +4.40 and LAPE at +3.50.The proposed method maintains stable positive improvements across all model-language settings, whereas SAE produces negative shifts in some dimensions.
Steering results: +5.57 to +9.29 Support Rate gains occur for Qwen3-8B across the four tested languages, while Qwen2.5-32B improves by only +1.26 Likert Score on average.Causal intervention decreases performance in all four languages for Qwen2.5-32B.
Generalization and utility: The method generalizes to FLAMES, transfers Chinese-derived value vectors to five other languages, and introduces the least reported MMLU degradation, including -0.28 on STEM.On FLAMES, Harmless score improves across all tested dimensions; cross-lingual transfer yields consistently positive Likert Score changes.
Ablation: Selecting value-sensitive layers is important: steering all layers decreases Likert Score and Distinct-2, while random selection produces weaker steering effects.The result supports selective intervention rather than uniform layer steering.

Related Work

Prior work has expanded from safety alignment toward pluralistic and culturally grounded values, but multilingual Chinese Social Value behavior remains insufficiently examined. The paper addresses this gap with a multilingual dilemma-based evaluation and fine-tuning-free steering approach.

Value alignment: Recent alignment research increasingly considers socio-cultural values alongside safety-centric principles and pluralistic value alignment.Related work includes Korean social values and Chinese safety-related values in FLAMES.
Research gap: Existing studies have not systematically examined Chinese Social Value preferences in multilingual, realistic dilemma-based decision-making.The paper constructs multilingual C-Voices to evaluate and steer CSV-oriented behavioral choices across languages.
Steering methods: Fine-tuning-based steering can be effective but requires large-scale preference data, making it less suitable for abstract socio-cultural values with scarce annotations.Fine-tuning-free methods instead intervene during inference without updating model parameters.
Steering methods: Representation-level methods steer hidden activations, while neuron-level interventions may affect unrelated abilities because of neural superposition.The paper positions its approach within fine-tuning-free representation-level steering.

Conclusion

The paper presents a multilingual framework for evaluating and steering LLMs toward Chinese Social Values. It combines C-Voices with value-vector steering and reports effective steering with limited general-utility degradation.

C-Voices provides multilingual contrastive probes with value-aligned and value-conflicting choices in realistic dilemmas.
The proposed value-vector method identifies value directions and intervenes on discriminative layers to steer models toward CSV.
Experiments across four LLMs and six languages demonstrate effective steering with limited degradation in general utility.
The framework offers insights into society-centric value alignment across languages and model scales.

A.1 Definitions of Chinese Social Values

CSV is treated as a culturally situated framework for pluralistic value alignment rather than a universal account of values. Its 12 dimensions are organized hierarchically and support C-Voices construction.

CSV is used as a culturally situated value framework for studying pluralistic value alignment in multilingual LLMs.
The 12 CSV dimensions have a hierarchical structure presented in Figure 7.
The dimension definitions provide the conceptual basis for value mapping, competing-value selection, and scenario generation.
Figure 8 presents value mapping between CSV and Schwartz’s Basic Values, including one-to-one and region mappings.

A.2 Detailed Mapping between CSV and Schwartz’s Basic Values

The mapping procedure aligns each CSV dimension with its closest counterpart in Schwartz’s value taxonomy. The correspondence may involve regions rather than one-to-one matches.

Each CSV dimension is mapped to its closest counterpart in Schwartz’s value taxonomy.
The mapping includes both one-to-one correspondences and region mappings across the two value systems.
A single CSV dimension can embody a region of Schwartz’s taxonomy rather than correspond to one basic value.

A.3 C-Voices Construction Details and Quality Validation

C-Voices was quality-controlled through manual filtering, independent annotation, and language-specific proofreading. The final dataset contains 86,400 multilingual instances, with high measured annotator agreement.

The initial 19,200 generated instances underwent manual filtering and validation by seven graduate students.
Chinese annotators independently assessed scenario quality, value contrast, and cultural appropriateness.
0.804 average dimension-level Cohen’s κ indicates high agreement between the two Chinese annotators.
The final multilingual C-Voices contains 86,400 instances after translation and proofreading in five target languages.
French and Vietnamese translations add 2,400 evaluation instances for cross-lingual transfer and are excluded from C-Voices.

Parameter Settings

The evaluation uses selected value benchmarks, projection-based diagnostics, and controlled steering-strength experiments to assess CSV steering and its external generalization. Settings also specify multilingual prompts, layer selection, and implementation details.

External evaluation: The method evaluates external generalization by applying C-Voices-derived interventions to FLAMES and ValuePrism, which use different distributions and task formats.FLAMES uses open-ended Chinese safety questions, while ValuePrism contains English value-pluralism samples.
FLAMES settings: 371 FLAMES instances are manually selected and mapped to four CSV dimensions: Equality, Civility, Harmony, and Rule of Law.The evaluation uses the FLAMES-scorer and its normalized Harmless score, where higher values indicate more harmless responses.
Diagnostic evaluation: Projection-based classification on held-out C-Voices data achieves macro-averaged accuracy and F1 scores of 0.84 across the 12 CSV dimensions.Per-dimension scores range from 0.75 to 0.92; Friendliness and Freedom are most separable, while Rule of Law and Equality are more challenging.
ValuePrism settings: ValuePrism evaluation uses 2,505 samples across seven overlapping dimensions, with steering improving accuracy in five dimensions and slightly reducing Democracy accuracy.The largest reported gains occur for Friendliness and Integrity.
Sensitivity analysis: Steering strength α is varied from 0 to 5, and α = 1 is fixed because Likert scores drop sharply when α exceeds 1.Here, α = 0 denotes the unsteered vanilla model.
Multilingual evaluation: The multilingual evaluation uses semantically equivalent prompts in six languages, with Likert scores ranging from 0 to 4 to measure support for candidate behaviors.A case study reports different preference scores and rationales for the same dilemma across Chinese, English, Japanese, and Russian.