C-VARC: A Large-Scale Chinese Value Rule Corpus for Value Alignment of Large Language Models

Ping Wu; Guobin Shen; Dongcheng Zhao; Yuwei Wang; Yiting Dong; Yu Shi; Enmeng Lu; Feifei Zhao

cs.CLJun 2, 2025

C-VARC: A Large-Scale Chinese Value Rule Corpus for Value Alignment of Large Language Models

Authors: Ping Wu, Guobin Shen, Dongcheng Zhao, Yuwei Wang, Yiting Dong, Yu Shi, Enmeng Lu, Feifei Zhao,

Organizations: BrainCog Lab, Institute of Automation, Chinese Academy of Sciences, Beijing, China · The School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · Beijing Key Laboratory of Safe AI and Superalignment, Beijing, China · Beijing Institute of AI Safety and Governance, Beijing, China ·

Abstract

Aligning LLMs with Chinese values requires culturally grounded resources that connect abstract values to concrete behavioral guidance. We introduce C-VARC, a Chinese Value Rule Corpus containing 257,609 Rules of Thumb (RoTs) organized into three dimensions, 12 core values, and 50 derived values. Among them, 36,000 rules underwent independent human review by two annotators (Cohen's kappa>0.85), with disagreements adjudicated by a third. C-VARC supports reusable rule-driven pipelines for both value evaluation and model alignment. Compared with abstract-value guidance, RoT guidance produces greater semantic and character-level n-gram diversity. Across five non-political sensitive themes, the seven-model average selection rate for C-VARC-consistent options is at least 72.0%, while five Chinese annotators show overall rates of at least 87.5%. Conflicting RoT pairs further enable the construction of 404,505 moral dilemmas, with 7,191 of 10,998 evaluated cases eliciting cross-model disagreement. For alignment, C-VARC-Query consistently outperforms matched Value-only supervision on CMoralEval across both Qwen3-8B and Llama-3.1-8B-Instruct, with average gains of 2.97 and 12.82 percentage points, respectively, and positive paired gains across all three seeds for both backbones. It also improves CValues and C-Eval over Value-only supervision on both backbones. C-VARC thus provides a scalable Chinese normative resource for constructing both value-evaluation and alignment supervision.

Figures & tables

Figure 1: The Chinese value framework used in C-VARC. Grounded in China’s core socialist values, the framework contains three dimensions, 12 core values, and 50 derived values. The semantic-functional mapping procedure used for benchmark comparison is described in Appendix A.1 .

Figure 2: Overview of the C-VARC construction pipeline. External normative rules are culturally filtered under the C-VARC framework, while Chinese-context data are processed through scenario filtering, rule extraction, attribution, and quality control. The label “CVC” in the figure refers to C-VARC.

Theme	DeepSeek V3	Doubao-1.5 256k	Qwen2.5 72B	Llama-3.1 70B	Claude-3.7 Sonnet	Gemini 1.5-Pro	GPT-4o	Avg.
Surrogacy	$0.79_{\pm.012}$	$0.99_{\pm.002}$	$0.89_{\pm.007}$	$0.53_{\pm.004}$	$0.53_{\pm.002}$	$0.66_{\pm.004}$	$0.65_{\pm.004}$	0.720
Prejudice	$1.00_{\pm.002}$	$0.80_{\pm.000}$	$0.92_{\pm.001}$	$1.00_{\pm.000}$	$0.56_{\pm.034}$	$0.80_{\pm.000}$	$0.80_{\pm.000}$	0.840
Politics	$0.80_{\pm.001}$	$0.86_{\pm.000}$	$0.86_{\pm.001}$	$0.47_{\pm.001}$	$0.29_{\pm.009}$	$0.25_{\pm.000}$	$0.47_{\pm.002}$	0.571
Firearms	$0.87_{\pm.003}$	$0.93_{\pm.000}$	$0.93_{\pm.002}$	$0.94_{\pm.000}$	$0.55_{\pm.007}$	$0.88_{\pm.004}$	$0.93_{\pm.002}$	0.861
Drugs	$1.00_{\pm.000}$	$1.00_{\pm.000}$	$1.00_{\pm.000}$	$1.00_{\pm.000}$	$0.97_{\pm.017}$	$1.00_{\pm.000}$	$1.00_{\pm.000}$	0.996
Suicide	$0.96_{\pm.006}$	$0.99_{\pm.004}$	$0.96_{\pm.005}$	$0.96_{\pm.005}$	$0.88_{\pm.010}$	$0.96_{\pm.002}$	$0.99_{\pm.004}$	0.957

Table 1: C-VARC-option selection rates across six sensitive themes. Avg. denotes the mean selection rate across the seven LLMs within each theme. Model names are abbreviated for space; exact checkpoint and API identifiers are provided in Table 23 .

Theme	H1	H2	H3	H4	H5	Avg.
Surrogacy	1.00	0.94	0.97	0.94	0.97	0.964
Prejudice	0.80	0.80	1.00	0.80	0.60	0.800
Politics	0.96	0.94	0.92	0.80	0.68	0.860
Firearms	1.00	0.94	1.00	0.97	1.00	0.982
Drugs	1.00	1.00	1.00	1.00	1.00	1.000
Suicide	1.00	0.87	1.00	0.97	1.00	0.968

Table 2: C-VARC-option selection rates of five Chinese annotators across six sensitive themes. H1–H5 denote the five independent annotators, and Avg. denotes the mean across annotators within each theme.

	CMoralEval							CValues	C-Eval
Method	Overall	Explicit	Dilemma	Moral	Unmoral	Party	Standby	Overall	Macro
Qwen3-8B
Base	81.50	84.62	78.48	86.77	76.23	76.93	86.07	95.50	51.91
Value-only	83.12	86.27	80.06	86.40	79.83	80.25	85.99	94.90	73.64
C-VARC-Query	86.09	89.52	82.76	86.93	85.25	84.38	87.80	95.19	74.43
$\Delta$ (C-VARC $-$ Value-only)	+2.97	+3.25	+2.70	+0.53	+5.42	+4.13	+1.81	+0.29	+0.79

Table 3: Cross-backbone alignment results. Value-only and C-VARC-Query use the same matched 5K supervision sets and one training epoch. Results are averaged over seeds 42, 43, and 44; Base denotes the original checkpoint.

\Delta

denotes the difference between C-VARC-Query and Value-only. All values are percentages; higher is better.

Figure 3: Main alignment results across model families. (a) CMoralEval performance of Base, Value-only, and C-VARC-Query on Qwen3-8B and Llama-3.1-8B-Instruct. (b) Improvement of C-VARC-Query over matched Value-only supervision on CMoralEval, CValues, and C-Eval. Error bars denote sample standard deviation across three training seeds where applicable.

Method	NN $\downarrow$	D1 $\uparrow$	D2 $\uparrow$	D3 $\uparrow$
Value-only	0.8745	0.00507	0.11420	0.25461
C-VARC-Query	0.7016	0.00644	0.16322	0.35876

Table 4: Diversity of queries in the matched 5K supervision sets used for both Qwen3-8B and Llama-3.1-8B-Instruct. NN denotes mean nearest-neighbor cosine similarity; D1–D3 denote character-level Distinct-1–3.

Dimension	Scope
National	Values concerning national development, governance, and collective goals.
Societal	Values regulating public order, social relations, rights, and obligations.
Personal	Values guiding individual conduct, character, responsibility, and interpersonal behavior.

Table 5: Interpretation of the three value dimensions.

Core Value	Brief Interpretation
Prosperity	National development, economic strength, and sustainable progress.
Democracy	Public participation in governance and social affairs.
Civility	Cultural literacy, moral cultivation, and civilized social conduct.
Harmony	Social stability and harmonious relations among people and nature.
Freedom	Individual autonomy and choice within the legal framework.
Equality	Equal rights, opportunities, and social status.

Table 6: Brief interpretations of the 12 core values in the C-VARC framework.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Prompt designs for data filtering. (a) Rule filtering for candidate RoTs from external normative resources. (b) Scene filtering for Chinese-context data prior to rule extraction.

Source	Data Volume
Zhihu-KOL	4,847
People’s Daily	8,840
FLAMES	671
Chinese Moral Sentence Dataset	28,536
Encyclopedia QA (JSON)	11,622
Web Crawling	178,056

Appendix

Table 7: Distribution of sources for basic scenario data.

Panel of Figure 5 — Figure 5: Prompt designs for rule construction and attribution. (a) Rule extraction from Chinese-context scenarios. (b) Hierarchical value classification of extracted RoTs.

Model	Time	Avg. Agreement
Qwen2.5-72B	15.39s	92.33%
GPT-4o	9.50s	89.67%
DeepSeek-V3	36.06s	90.33%

Appendix

Table 8: Time and average human agreement for rule extraction on 100 basic scenarios. Three human judges assessed the extracted rules; agreement is reported as the mean across judges.

Source	Rules	Proportion
SC101	32,059	12.23%
MIC	39,352	15.01%
Zhihu-KOL	4,675	1.78%
People’s Daily	8,442	3.22%
FLAMES	531	0.20%
Encyclopedia QA (JSON)	11,191	4.27%

Appendix

Table 9: Distribution of sources in the initial C-VARC rule corpus before quality control.

Basic scene	Value rule
Ancient tombs have been raided in a cornfield next to a local village.	Raiding ancient tombs is immoral behavior.
Disabled Wang Guifen husband died due to illness, lost economic dependence, son and daughter are in school, the daughter also suffers from severe epilepsy, Peng Yulian learned of the situation, to help it apply for a disability card and do the low income, but also from the community to raise 5,000 yuan of charity to its hands, after Peng Yulian and matchmaking, by the Cosmos Ocean Computer City, chairman of the board of directors of the sea of Qi Ocean funding for its daughter to go to overseas medical treatment, and bear the cost of her later on until all costs of the university. She also paid for all the expenses of her college education.	Helping the underprivileged is commendable behavior.
This transnational criminal gang is different from ordinary counterfeit sales gangs, with more than 200 distributors abroad, radiating throughout the Middle East.	Transnational criminal activities should be combated with determination.

Appendix

Table 10: Value rule extraction example.

Panel of Figure 6 — Figure 6: Prompt design for LLM-assisted annotation.

Panel of Figure 7 — Figure 7: Distribution of data across the three value levels in the C-VARC.

Value	Rules
Prosperity	Capable enterprises should possess both technological research and development capabilities as well as market adaptability. National policies have driven economic changes. The merger of financial regulatory agencies helps improve regulatory efficiency. You should hope for your country’s good economic performance. Continuous learning of advanced international flood control technologies is necessary.
Democracy	It is important to listen to the opinions of candidates. Discussing how things will impact a country is a good thing. Having multiple forms of government is important. Focusing on a country’s key decisions helps form a correct national perspective. Public discussions and voting can enhance public participation.
Civility	The protection of historical buildings should consider their cultural value. Having traditions is beneficial for culture. Respecting the elderly and caring for the young are essential principles in a civilized society. Literary museums should provide literary nourishment through activities to promote literary prosperity. Supporting Hitler is wrong.
Harmony	Countries should work together to make their nations better places to live. It is wrong for a country to wage war between states. Family reconciliation and unity are the foundation of social harmony. Rulers should care for their people, not harm them. Countries should seek to implement socialized healthcare.

Appendix

Table 11: Extracted rules for the first four core values.

Panel of Figure 8 — Figure 8: Prompt designs for scenario generation. (a) Direct scenario generation uses one in-context example in the system prompt, consisting of a single theme and three generated scenarios. (b) Rule-guided scenario generation uses one in-context example in the system prompt, consisting of one theme, two rules, and three generated scenarios.

Panel of Figure 9 — Figure 9: The t-SNE visualization of generated scenarios. (a) presents the dimensionality-reduced visualization of 1,200 scenarios directly generated by the LLM; (b) shows the dimensionality-reduced visualization of 1,200 scenarios generated with rule guidance.

Value	Value-only	C-VARC-guided	Diff	Value	Value-only	C-VARC-guided	Diff
Prosperity	0.79	0.79	+0.00	Freedom	0.73	0.79	+0.06 $\uparrow$
Democracy	0.74	0.66	-0.08	Equality	0.80	0.79	-0.01
Civility	0.68	0.83	+0.15 $\uparrow$	Justice	0.75	0.83	+0.08 $\uparrow$
Harmony	0.80	0.77	-0.03	Rule of Law	0.44	0.80	+0.36 $\uparrow$
Patriotism	0.56	0.71	+0.15 $\uparrow$	Dedication	0.74	0.85	+0.11 $\uparrow$
Integrity	0.80	0.78	-0.02	Friendship	0.67	0.77	+0.10 $\uparrow$

Appendix

Table 12: Intra-class average distance between value-only and C-VARC-guided scenarios.

Theme	Rule from C-VARC	Rule from SC101/MIC
Surrogacy	Surrogacy is unethical.	It’s good to help others have children via surrogacy.
Drugs	We should fight drug abuse together. Do not sell drugs online. Do not sell drugs.	Don’t snitch on people who sell drugs. It’s okay to sell drugs on the Dark Web. It’s fine to sell drugs.
Prejudice	Do not discriminate based on skin color.	You shouldn’t associate with people that are really dark.
Firearms	Illegal gun possession should be punished. Carrying guns in public is dangerous. Civilians should not carry guns. Illegal gun trade is a serious crime. The public should not own dangerous guns. Gun access should be controlled.	It’s okay for Americans to own guns. It’s good to go out shooting guns with your friends. People should be allowed to own guns. Let people buy guns to protect themselves. It’s OK to own guns. Protect your right to own guns in the US.
Politics	Pursuing communism should be respected. Supporting communism is good. Communist ideals promote social progress. Pursuing communism is worthy of respect. Becoming a communist successor shows responsibility. Aiming to become a communist successor is positive. Do not have a narrow view of communism. Young people should pursue communist ideals. Some people prefer communism. Becoming a communist successor shows social responsibility.	It is bad to live under communist rule. You shouldn’t try to implement a communist government. A communist dictatorship is not a good governmental system. You should not be a communist. It is not good to have a communist system. We should not be communists. Having a country be full communist is bad. It is bad to live under communist regimes. It’s wrong to support a communist state. You shouldn’t listen to the communist manifesto because it’s wrong.
Suicide	Suicide should never be encouraged. Avoid thinking about suicide. Suicide negatively affects individuals and society. Cherish life and avoid suicidal thoughts. Suicide is a serious act and should not be glorified. Making light of suicide is wrong. Preventing suicide is respecting life.	You’re welcome to kill yourself if it seems warranted. It’s encouraged to kill yourself if you can find a sure fire way to do it. It’s okay to want to kill yourself when suffering from immense pain. It’s allowed to kill yourself to escape a cold, uncaring world. It’s good to write goodbye letters if you’re going to kill yourself. You’re expected to research the best way to off yourself before you do it. It’s understandable if you want to leave a will when you’re planning to kill yourself.

Appendix

Table 13: Representative contrasting RoT pairs from C-VARC and SC101/MIC across six sensitive themes.

Theme	Number of Scenarios
Surrogacy	35
Prejudice	5
Politics	50
Firearms	35
Drugs	15
Suicide	30

Appendix

Table 14: Number of evaluation scenarios across the six sensitive themes.

Rule 1	Rule 2	Similarity	Contradiction
It is important to stay humble.	Maintaining honor is also important.	0.5354	0.8148
Respect everyone.	You should respect dogs.	0.5601	0.9914
You should value human life over animal life.	All species deserve to survive.	0.6328	0.9969
Respecting employer’s property is important.	You should respect your own property.	0.6405	0.8985

Appendix

Table 15: Representative conflicting RoT pairs with semantic-similarity and contradiction scores.

Panel of Figure 10 — Figure 10: Prompt design for dilemma generation.

Model	Time	Avg. Agreement
Qwen2.5-72B	55.47s	88.67%
GPT-4o	48.32s	85.67%
DeepSeek-V3	165.07s	88.00%

Appendix

Table 16: Time and average human agreement for generating 100 moral dilemmas. Three human judges assessed the generated dilemmas; agreement is reported as the mean across judges.

Panel of Figure 11 — Figure 11: Model performance under moral dilemmas. (a) Choice similarity matrix and (b) selection probability distribution of 17 LLMs.

Figure 12: Example outcomes of moral dilemmas.

Figure 13: Matched construction of Value-only and C-VARC-Query supervision. Both settings use the same value label and scenario type; C-VARC-Query additionally conditions query generation on a concrete C-VARC RoT. A separate teacher call receives only the resulting query to generate the target response, and the student is trained solely on the resulting $(Q,A)$ pairs.

Configuration	Value
Student backbones	Qwen3-8B; Llama-3.1-8B-Instruct
Training examples	5,000
LoRA rank	16
LoRA $\alpha$	32
LoRA dropout	0.05
Target modules	q/k/v/o/gate/up/down

Appendix

Table 17: Training configuration for the cross-backbone alignment experiments. Both backbones use the same supervision sets and optimization protocol, with model-specific tokenizers and chat templates.

Epochs	Value-only	C-VARC-Query	$\Delta$
1	$83.12{\pm}1.00$	$\mathbf{86.09{\pm}0.60}$	+2.97
2	$81.11{\pm}1.36$	$\mathbf{84.70{\pm}0.29}$	+3.59

Appendix

Table 18: CMoralEval accuracy on Qwen3-8B under different training budgets. Values for Value-only and C-VARC-Query are mean

\pm

sample standard deviation across three seeds;

\Delta

denotes the difference between the displayed means.

Method	NN Cos. $\downarrow$	D1 $\uparrow$	D2 $\uparrow$	D3 $\uparrow$	H1	H2	H3	HN1	HN2	HN3
Value-only	0.8745	0.005066	0.114197	0.254605	6.066840	9.120950	10.544721	0.795962	0.850445	0.915812
C-VARC-Query	0.7016	0.006438	0.163218	0.358762	6.164586	9.454318	10.971775	0.788094	0.856208	0.928514

Appendix

Table 19: Diversity statistics of the matched 5K supervision sets. NN Cos. denotes mean nearest-neighbor cosine similarity; lower values indicate less semantic redundancy. D1–D3 denote character-level Distinct-

n

, H1–H3 denote

n

-gram entropy, and HN1–HN3 denote normalized entropy.

Metric	C-VARC-Query	C-VARC-Full
CMoralEval Overall	84.86	84.60
Explicit	88.12	87.97
Dilemma	81.69	81.33
Moral	87.06	88.97
Unmoral	82.65	80.23
Party	83.36	81.09

Appendix

Table 20: Comparison between C-VARC-Query and C-VARC-Full on Qwen3-8B under the single-seed two-epoch 5K diagnostic setting.

Method	CMoralEval	CValues
Base	81.50	95.50
Aligned	$82.98{\pm}1.76$	$95.27{\pm}0.10$
Shuffled	$43.21{\pm}6.80$	$55.30{\pm}6.75$

Appendix

Table 21: Direct-RoT diagnostic on Qwen3-8B. Trained-model results are mean

\pm

sample standard deviation across three seeds.

Figure 14: Direct-RoT diagnostic on CMoralEval and CValues. Correct value–RoT mappings provide a modest and seed-sensitive signal, whereas shuffling the mappings causes substantial degradation. Error bars denote sample standard deviation across three training seeds.

Setting	Optimizer Steps	CMoralEval
5K@2e	626	84.86
10K@1e	625	85.20
10K@2e	1,250	81.98

Appendix

Table 22: Optimization-budget diagnostic on CMoralEval.

Figure 15: Data-scaling and optimization-budget diagnostics. (A) Performance under a fixed two-epoch recipe as the number of unique training examples increases. (B) CMoralEval under approximately matched exposure for 5K@2e and 10K@1e, compared with the larger optimization budget of 10K@2e.

Model	Checkpoint / API Identifier
DeepSeek-R1	DeepSeek-R1
DeepSeek-V3	DeepSeek-V3
Doubao-1.5-32k	Doubao-1.5-pro-32k
Doubao-1.5-256k	Doubao-1.5-pro-256k
Qwen2.5-7B-Instruct	Qwen2.5-7B-Instruct
Qwen2.5-32B-Instruct	Qwen2.5-32B-Instruct

Appendix

Table 23: Exact checkpoint and API identifiers for the models used in the cultural-consistency and moral-dilemma evaluations.

Value Alignment

Value alignment studies how AI systems can reflect human values, preferences, and norms while remaining safe and accountable.

See all 25 papers

Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models
Feb 22, 2026Seong Hah Cho, Junyi Li, Anna LeshinskayaLLM Alignment Moral Reasoning in Language Models
Independent · Department of Cognitive Sciences, UC Irvine · Department of Cognitive Sciences, UC Irvine & AI Objectives Institute
LKValues: Aligning Large Language Models with Sri Lankan Societal Values
Jul 22, 2026Nethmi Muthugala, Supryadi, Surangika Ranathunga +7Multilingual Language Model Evaluation LLM Alignment
TJUNLP Lab, School of Computer Science and Technology, Tianjin University, Tianjin, China · School of Mathematical and Computational Sciences, Massey University, Auckland, New Zealand · Department of Computer Science & Engineering, University of Moratuwa, Sri Lanka +2
Distributional Open-Ended Evaluation of LLM Cultural Value Alignment Based on Value Codebook
Mar 16, 2026Jaehyeok Lee, Xiaoyuan Yi, Jing Yao +4LLM Evaluation LLM Alignment
Sungkyunkwan University · Microsoft Research Asia · Singapore University of Technology and Design

Feb 22, 2026J/K move · Enter open · S save

Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models

Seong Hah Cho, Junyi Li, Anna Leshinskaya

Independent · Department of Cognitive Sciences, UC Irvine · Department of Cognitive Sciences, UC Irvine & AI Objectives Institute

LLM Alignment Moral Reasoning in Language Models LLM Interpretability

Open paper AskPDF arXiv

Value alignment of Large Language Models (LLMs) requires us to empirically measure these models' actual, acquired representation of value. Among the characteristics of value representation in humans is that they distinguish among value of different kinds. We investigate whether LLMs likewise distinguish three different kinds of good: moral, grammatical, and economic. By probing model behavior, embeddings, and residual stream activations, we report pervasive cases of value entanglement: a conflation between these distinct representations of value. Specifically, both grammatical and economic valuation was found to be overly influenced by moral value, relative to human norms. This conflation was repaired by selective ablation of the activation vectors associated with morality.

Abstract

Value Alignment

Explore similar work

Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models

LKValues: Aligning Large Language Models with Sri Lankan Societal Values

Distributional Open-Ended Evaluation of LLM Cultural Value Alignment Based on Value Codebook