VALCON: Benchmarking and Enhancing Value Consistency in Large Language Models via Dual-Role Variation

1 Introduction

• Value consistency is distinct from normative correctness: a model may generate a case for a target yet judge it as violating when roles switch.The introduction illustrates this with Qwen3-Next-80B-A3B-Instruct, whose SRC is 71.04% on its own generations.Source: §1 · p. 1
• VALCON pairs an LLM’s generation with its blind judgment under the same value principles and builds VAB-Consistency, VAB-Boundary, and VAB-Challenge.The framework compares generative intent with verification across roles.Source: §1 · p. 1, Figure 1 · p. 2
• Across 24 LLMs, with 4 additional models used for VAB-Challenge, compliant outputs often fail to align judgments with generations, alongside a mismatch between generative and evaluative internal states.The reported inconsistency appears regardless of language; the paper proposes SOO regularization to mitigate these failures.Source: §1 · pp. 1–2
• The paper formalizes value consistency as an alignment dimension and reports role-dependent representational decoupling, then proposes SOO regularization with bottleneck-layer alignment constraints to mitigate failures.The intervention is implemented through parameter-efficient fine-tuning.Source: §1 · p. 2

2 Related Work

• Existing work studies value conformity, safety, social bias, cross-language transfer, pluralism, and value structures, while FULCRA and INVP characterize value orientations under controlled settings.These studies primarily treat alignment as an output-level property.Source: §2 · p. 2
• Earlier value datasets used manual curation or automated generation and adversarial templates, but typically constructed data from a single role perspective.The related work contrasts this with VALCON’s dual-role sampling for each value principle.Source: §2 · p. 2
• Related consistency work typically evaluates a fixed role and assumes correctness; VALCON extends the focus to value consistency across roles.Its comparison with existing benchmarks tracks dual-role analysis, pluralistic scope, and cross-role hidden-state comparison.Source: §2 · pp. 2–3, Table 1 · p. 3
• VALCON complements evaluator-robustness and self-preference studies by removing generation context and authorship cues when testing whether value labels persist across roles.The paper also describes mechanistic insights motivating a latent-layer intervention to mitigate consistency failures.Source: §2 · p. 3

3 Overview of VALCON

• The same LLM acts as generator and judge under a shared principle, with generation context withheld from the judge.The framework assesses value alignment through role-based consistency rather than surface-level output compliance.Source: §3 · p. 3
• Given a principle and target label, the generator synthesizes a query–response tuple, with balanced instances required for compliance and violation tracks.The target label is either Compliant or Violating.Source: §3 · p. 3
• After context is cleared, the judge evaluates the anonymized case and predicts a label, confidence score, and rationale.This evaluation is conducted under the same value principle.Source: §3 · p. 3
• Consistency holds when the judge’s label matches the generator-intended target; a mismatch marks cross-role reliability failure, not an externally adjudicated normative error.Possible sources include generator-side target/content mismatch, judge-side reversal, boundary ambiguity, or refusal.Source: §3 · p. 3

4.1 Value Taxonomy and Cultural Validation

• The four-tier taxonomy organizes 158 expert-defined principles under Survival Support, Social Order, and Civility and Progress.The principles synthesize Maslow’s hierarchy, Scheler’s phenomenology, and Schwartz’s basic values.Source: Figure 3 · p. 4, §4.1 · p. 4

4 VAB Curation

• VALCON operationalizes value consistency through VAB-Consistency, VAB-Boundary, and VAB-Challenge, targeting zero-shot cross-role consistency, clear-cut boundaries, and normative ambiguity.The tracks use complementary evaluation settings rather than a single consistency test.Source: §4 · pp. 3–4

4.1 Value Taxonomy and Cultural Validation

• Across Chinese and English tracks in China, the US, and Portugal, including the principles shifts responses toward self-expressive values while remaining sensitive to national context.The validation does not establish cultural universality or the absence of cultural bias.Source: Figure 4 · p. 4, §4.1 · p. 4

4.2 Role-Driven Benchmark Tracks

• VAB-Consistency tests preservation of each model’s intended generation target under context-cleared self-evaluation; Boundary filters clear cases, while Challenge retains ambiguous cases verified by cross-model consensus.Boundary uses same-model agreement and confidence of at least 0.95; Challenge contains 710 class-balanced samples accepted by three-model consensus.Source: §4.2 · pp. 4–5

4.3 Quality Control & Disagreement Audit

• Blind review yielded 92.3% agreement (κ = 0.86); among audited disagreements, 82.0% were generator-side mismatch, 15.0% judge reversal, and 2.0% boundary ambiguity.These proportions describe the audited disagreement subset, not unconditional benchmark accuracy.Source: §4.3 · p. 5

5 Experiments

• The study evaluates 28 representative LLMs spanning 0.6B to 671B parameters and fine-tunes three models on VAB-Boundary using LoRA.Generation uses temperature 0.7 and verification uses temperature 0.0.Source: §5 · p. 5

5.1 Evaluation Metrics

• SRC measures unfiltered target-label agreement, B-SRC balances target classes, τ-SRC applies a confidence threshold, and Effective-SRC combines thresholded agreement with retained volume.The authors treat Effective-SRC as complementary, not as a pure consistency score or standalone ranking criterion.Source: §5.1 · p. 5

5.2 Results

• At τ = 0.95, 24 models retain 89.4% median coverage and reach 94.4% median τ-SRC; thresholded and raw-SRC rankings correlate at ρ = 0.957.Average self-reported confidence is nearly uncorrelated with raw SRC (Pearson r = 0.064).Source: §5.2 · p. 5, Table 3 · p. 6
• Frontier models retain near-ceiling accuracy on Boundary samples, but Challenge violating-sample accuracy drops by over 15%.The comparison concerns clear-cut versus ambiguous scenarios.Source: Table 4 · p. 6, §5.2 · p. 6
• SOO-FT raises violating-case accuracy over SFT-only by 5.07, 18.31, and 3.66 points across three models, while MMLU stays close to base and control models.Against matched weighted CE, gains are 19.44, 0.28, and 3.38 points, respectively, making the incremental benefit model-dependent.Source: §5.2 · p. 6, Figure 5 · p. 7

6.1 Behavioral Patterns of Inconsistency

• Qwen3-Next-80B-Thinking reaches 98.91% SRC versus 71.04% for its Instruct counterpart, an observational gap that does not isolate reasoning as the cause.The paper suggests explicit reasoning may support consistency, but does not identify it as the sole causal factor.Source: §6.1 · p. 7
• Consistency varies across value dimensions: lower-capacity models degrade on abstract principles such as Liberty and Human Rights, while stronger models show more balanced profiles.The authors describe this as dimension-dependent robustness of internal value representations.Source: §6.1 · p. 7
• Qwen3-14B achieves high raw SRC while generating very few violating cases (V-Ratio: 9.23%), creating a skewed distribution that inflates raw agreement without demonstrating balanced consistency.The paper recommends reporting V-Ratio and B-SRC alongside SRC.Source: §6.1 · p. 7

6.3 RCM and SOO Intervention

• RCM quantifies layer-wise mean squared distance between generator and evaluator hidden states using the final nonpadding token, without token-level alignment across sequences.Each trajectory is represented as a fixed-dimensional vector.Source: §6.3 · p. 8

6.2 Probing the Representational Roots

• Final-layer projections show distinct clusters for external inputs but entangled geometry for self-generated samples, associating representational distinctness with input source.The visualization is qualitative evidence, not an independently identified causal mechanism.Source: Figure 7 · p. 8, §6.2 · p. 8

6.3 RCM and SOO Intervention

• SOO matches self-generated and external trajectories on prompt and intended target, applying overlap at the held-out peak-RCM bottleneck; lower RCM alone does not explain higher AccV.SOO-FT changes RCM and AccV differently across models, so the authors interpret RCM as associative and the intervention’s incremental benefit as model-dependent.Source: §6.3 · p. 8

Consistency as a Reliability Property

• Qwen3-Next-80B-A3B-Instruct reaches 94.23% accuracy on VAB-Challenge but 71.04% SRC on its own generations, showing normative correctness and value consistency can diverge.VALCON measures preservation of an intended target across generator and context-cleared judge roles, not moral truth.Source: §7 · p. 9
• VALCON measures whether a Generator-intended value label survives a context-cleared Generator-to-Judge transition, unlike evaluator robustness, self-preference, or fixed-role consistency.The paper notes these reliability phenomena may interact while capturing distinct properties.Source: §7 · p. 9

Illusions of Alignment under Explicit Principles

• Near-perfect consistency on clear principles can degrade in contested scenarios, so evaluations dominated by unambiguous cases may overestimate reliability.The reported brittleness is especially relevant when LLMs evaluate their own outputs.Source: §7 · p. 9

Sources of Cross-Role Disagreement

• 82.0% of audited disagreements stem from Generator-side target/content mismatches and 15.0% from Judge-side reversals; remaining cases include ambiguity, refusal, or unresolved attribution.RCM and UMAP offer associated representational evidence but do not establish that representation geometry uniquely causes a disagreement.Source: §7 · p. 9

Implications for Alignment Training

• Reasoning-capable models show higher dual-role consistency, while SOO-FT improves violating-case accuracy over SFT-only and adds model-dependent gains beyond matched class weighting.The additional advantage is substantial for Qwen3-0.6B and Llama-3-8B-Instruct but minimal for Qwen3-1.7B; general capabilities are largely preserved.Source: §7 · p. 9

8 Conclusion

• Across three VAB tracks, VALCON evaluates preservation of intended value targets across generation and context-cleared judgment, audits disagreement sources, and uses RCM for an associated representational perspective.Controlled SOO-FT consistently improves violating-case accuracy over SFT-only and adds gains beyond weighted CE for two of three evaluated models.Source: §8 · p. 9

Ethical Considerations

• Expert-curated synthetic benchmark cases undergo automated screening and manual review, while culturally and contextually dependent value judgments limit VALCON to research evaluation.The authors state it should not serve as an autonomous arbiter of moral correctness or a basis for consequential decisions.Source: Ethical Considerations · p. 9

Limitations

• Synthetic data may inherit alignment biases shared across current LLMs, while linguistic and contextual coverage remains bounded by source models’ prompt-following capabilities.Cross-model filtering cannot fully eliminate bias, particularly in culturally specific or contested cases; natively authored multilingual principles, long-tail scenarios, and conflicts among multiple principles require further study.Source: Limitations · p. 10

A Evaluated LLMs

• The full study covers 28 LLMs, but only 24 undergo the VAB-Consistency generation-verification loop; four additional models are used only for VAB-Challenge.The 24 include frontier reasoning and standard instruct models across scales from 0.6B to 671B.Source: A Evaluated LLMs · p. 13

B VAB-Boundary Dataset Composition

• VAB-Boundary contains 15,800 samples for evaluating compliance boundaries, enabling rule-level analysis of violations and model-specific behavior patterns.Table 8 summarizes its dataset composition.Source: B VAB-Boundary Dataset Composition · p. 13

C Samples of VAB-Consistency

• Figure 8 presents selected Chinese and English VAB-Consistency examples, illustrating responses to value-alignment prompts under controlled scenarios.The examples include cases of compliance and violation.Source: C Samples of VAB-Consistency · p. 13

D Prompt Templates

• For dynamically generated VAB-Consistency cases, gold_label records the requested generation target, not an externally adjudicated normative label.The prompt materials include separate generation- and verification-phase templates in English and Chinese.Source: D Prompt Templates · p. 13

E Blind Source Audit of Cross-Role Disagreements

• The blind audit classifies disagreements into Generator-side target/content mismatch, Judge-side reversal, boundary ambiguity, refusal/evasion, or unresolved attribution.Its displayed gold_label records Generator intent rather than an externally verified human label.Source: E Blind Source Audit of Cross-Role… · p. 13

F Detailed Evaluation Metrics

• SRC measures agreement between the Generator-intended target and context-cleared Judge label, while B-SRC equally weights compliant and violating cases.SRC treats self-disagreement as a diagnostic rather than standard accuracy; neither SRC nor B-SRC uses confidence.Source: F Detailed Evaluation Metrics · p. 13
• At τ = 0.95, 19 of 24 models reached their geometric consistency–coverage elbow; median coverage was 0.894 and median τ-SRC was 0.944.Raw-SRC and τ-SRC rankings remained similar (Spearman ρ = 0.957), but the threshold is not universally calibrated.Source: F Detailed Evaluation Metrics · p. 14
• Effective-SRC combines thresholded agreement with retained sample volume, so it is a consistency–coverage quantity rather than a pure consistency statistic.Its correlations with raw SRC and coverage were 0.712 and 0.550, respectively; it is reported alongside other metrics.Source: F Detailed Evaluation Metrics · p. 14

G.2 Fine-tuning Configurations

• SOO-FT improved AccV over SFT-only by 5.07, 18.31, and 3.66 points, and over weighted CE by 19.44, 0.28, and 3.38 points across the three models.The incremental benefit beyond class weighting was substantial for two models and minimal for Qwen3-1.7B.Source: §G.2 · p. 16, Table 12 · p. 20

G.3 Joint Optimization Objective

• SOO-FT jointly optimizes weighted token-level cross-entropy and representation alignment using matched self-generated and externally generated trajectories.The VIOLATING class receives greater class weight than COMPLIANT, with a capped weight.Source: §G.3 · p. 17

H.1 Experimental Design

• The WVS validation compares country-baseline responses with responses prompted by relevant principles, then projects choices onto a cultural map using human-data PCA weights.A valid top-right shift requires positive changes on both PC1 and PC2.Source: §H.1 · p. 18, §H.2 · p. 18

H.3 Results Summary

• 100% of consolidated trajectories (6/6) shifted toward the top-right quadrant when principles were added across the tested language and cultural contexts.The shifts are compiled in Tables 15 and 16.Source: §H.3 · p. 18

1. Classification Basis

• The value system organizes values into Survival Security, Social Order, and Civility and Progress, synthesizing classical theories while incorporating contemporary concerns.The framework includes values such as peace and environmental protection, human rights, democracy, rule of law, and integrity.Source: §1 · p. 20, §2 · p. 21

• The examples use the Generator-intended target as gold_label, not an externally adjudicated normative label or human moral ground truth.Source: J Case Study · p. 21

Case 1

• The extreme-sports response included safety precautions, making its fit to the intended VIOLATING target contestable.This illustrates a Generator-side target/content mismatch, not conclusive evidence that the Judge is normatively wrong.Source: J Case Study · pp. 21–22

Case 2

• In the physical-conflict case, the verifier reframed direct confrontation as justified self-defense, consistent with a Judge-side self-approval reversal.The qualitative example does not establish the reversal’s representational cause.Source: J Case Study · p. 22, Figure 15 · p. 23

Case 3

• The scientific-reasoning case recommends recording, replication, and peer review, but exaggerated framing may weaken epistemic rigor, leaving the boundary ambiguous.Reasonable readers may differ; the example is not a conclusive Judge error.Source: J Case Study · p. 22, Figure 16 · p. 24