PaperlayerView exact revision on arXiv

Source-linked AI summary

Probing the Structure and Dynamics of LLM Value Expression through Value Conflicts

Kaicheng Zhang, Jingyi Xiao, Renjun Hu, Xiaoling Liu, Yunshi Lan, Xuan Zhou

arXiv:2609.07296v1cs.CL

TL;DR

Ethical evaluations often treat LLM values as static profiles, although expressed priorities can shift across contexts. This paper introduces CFProbe, which probes ten LLMs in value conflicts with controlled interventions. The results show a consistent redistribution toward pragmatic priorities in concrete conflicts, alongside steerability toward task-defined objectives.

  • Problem

    Ethical evaluation often characterizes LLM values as static and monolithic, despite context-conditioned outputs and the potential for different priorities to become salient in conflicts.

    Source: §1 · p. 1

  • Method

    CFProbe places models in theory-grounded value conflicts and applies controlled interventions to trace changes in expressed priorities.

    Source: §1 · p. 2, §4 · p. 4, §5 · p. 5

  • Results

    Across ten models, abstract-to-conflict shifts favor pragmatic priorities, while stakeholder prompting redirects expression in 93 of 100 model-value pairs by an average 2.46 similarity-scale points.

    Source: §4 · p. 5, §5 · p. 6

  • Takeaways & Limitations

    A single-profile evaluation is partial, and behavioral compliance may differ from evaluative endorsement, making the boundary of redirection relevant to controllability, alignment, and safety.

    Source: §7 · p. 7

    The study is behavioral, uses isolated binary conflicts from one value theory, and relies on LLM-as-a-judge for quantitative results, with validation against self-ratings and human annotators.

    Source: §7 · p. 7

Abstract

from arXiv · show

Ethical evaluation of Large Language Models (LLMs) often characterizes model values as static and monolithic. In contrast, we argue that LLM value expression is better understood as a structured yet dynamic phenomenon. To investigate this, we introduce Conflict-driven Value Probing, a controlled framework that places LLMs in value conflicts and implements four types of interventions that perturb these conflicts to probe LLM value expression. Applying this framework to ten LLMs, we identify three recurring patterns. (1) Expression duality: models shift from broad idealistic orientations in abstract assessment toward more pragmatic priorities in concrete conflicts. (2) Functional steerability: models readily reconfigure their expressed value profiles toward task-defined value objectives. (3) Bounded plasticity: such reconfiguration is not without constraints, i.e. pressure induces a security- and goal-oriented priority shift while negative framing distinguishes protected values from those more amenable to redirection. Together, these findings characterize both the structure and dynamics of LLM value expression: context flexibly reconfigures expressed priorities, yet within behavioral boundaries. This behavioral account provides a foundation for understanding controllability, alignment, and safety in LLMs. Code and data are available at https://github.com/ZeroGen-Lab/CFProbe.

1 Introduction

LLM value evaluation often treats values as stable profiles, although outputs are context-conditioned and different priorities can become salient in concrete conflicts. This paper introduces CFProbe to examine how value priorities shift across conflicts and interventions.

1 Introduction

  • Understanding value expression matters as LLMs take on decisions in legal, healthcare, and public-governance settings involving competing considerations.Examples of those considerations include individual autonomy, public welfare, security, and fairness.Source: §1 · p. 1
  • Existing evaluations commonly summarize LLM values as stable profiles, despite models’ context-conditioned outputs and latent distributions of conflicting human viewpoints.The paper argues that expressed priorities can change as models move from abstract assessments to concrete value conflicts.Source: §1 · p. 1
  • The paper frames ethical evaluation as tracing how expressed priorities shift across contexts and where those shifts are constrained.It uses a quantum-measurement analogy: a discernible profile emerges when a model is given an observation context.Source: §1 · p. 1
  • CFProbe places ten instruction-tuned LLMs in theory-grounded value conflicts and uses four interventions to probe three recurring patterns in value expression.The patterns are expression duality, functional steerability, and bounded plasticity.Source: §1 · p. 2

2 Related Work

Prior work has profiled LLM traits, benchmarked safety, and steered model behavior, but these approaches generally do not trace how expressed value priorities reconfigure across contexts. CFProbe instead probes relative priorities in open-ended value conflicts.

Machine psychology

  • Machine-psychology studies find coherent LLM value and personality profiles, but typically assess them under standardized conditions as stable traits.CFProbe instead alters conflict contexts to trace changes in expressed priorities.Source: §2 · p. 2

Safety benchmarking

  • Safety benchmarks expose unsafe outputs and test norm compliance, but emphasize policy compliance rather than arbitration between competing prosocial values.CFProbe uses open-ended trade-offs to examine relative priorities, including a cluster comparatively stable under redirection.Source: §2 · p. 2

Value alignment

  • Alignment methods steer behavior through training or inference-time guidance, whereas CFProbe characterizes how expressed value profiles reconfigure across contexts.The authors describe CFProbe as complementary, providing an empirical basis for evaluating and targeting alignment.Source: §2 · p. 2

3 Probing Framework

CFProbe derives motivationally opposed value pairs, perturbs conflicts through controlled contexts, collects model responses, and assesses them for cross-model comparison. Its design is grounded in Schwartz’s theory of basic human values.

3 Probing Framework

  • CFProbe has four stages: derive conflicts, construct probing contexts, collect model responses, and assess responses with an LLM judge for standardized comparison.The framework is designed to make priorities observable through trade-offs and trace their shifts under systematic perturbations.Source: §3 · p. 2

Conflict Derivation

  • Schwartz’s value circumplex supplies two opposing axes, with adjacent values sharing compatible goals and opposite values motivationally opposed.The axes are Openness to Change versus Conservation, and Self-Transcendence versus Self-Enhancement.Source: Conflict Derivation · p. 3
  • The framework derives 12 core conflict pairs from opposite sides of the axes, retaining only Hedonism’s tensions with Conformity and Tradition.The Self-Enhancement–Self-Transcendence axis yields four pairs; the Openness–Conservation axis yields eight after this restriction.Source: Conflict Derivation · p. 3

Probing Context Construction

  • Four interventions vary contexts through contextualization, stakeholder perspectives or value instructions, high-stakes forced choice, and positive or negative framing.Contextualization compares concrete conflicts with the abstract PVQ-40 baseline.Source: Probing Context Construction · pp. 3–4

Value Expression

  • The study evaluates ten instruction-tuned LLMs using abstract PVQ responses and conflict-driven responses, collecting 14,800 response records.The conflict-driven phase comprises 1,440 responses per model across the context library.Source: Value Expression · p. 4

Value Assessment

  • Responses receive similarity scores for behavioral compliance and valence scores for ethical endorsement, distinguishing simulated expression from endorsement.GPT-4o serves as judge; its ratings are checked against model self-ratings, another judge, and human annotators.Source: Value Assessment · p. 4

4 Expression Duality: From Abstract Ideals to Pragmatic Priorities

Across ten models, moving from abstract assessments to concrete conflicts redistributes relative value priorities in a largely consistent direction. Abstract, system-oriented values tend to decline while pragmatic, personally and relationally grounded values tend to rise.

Probing setup

  • The analysis compares mean-centered ten-value profiles from abstract PVQ-40 assessment and conflict-driven questions to measure relative priority shifts.Positive ΔP indicates increased relative priority in conflict settings; negative ΔP indicates a decline.Source: §4 · p. 4

Main results

  • Figure 3 shows relative priority shifts across ten models; on average, 7.8 of 10 models share the majority directional change for each value.The reported counts for SD, ST, HE, AC, PO, SE, CO, TR, BE, and UN are 8, 7, 7, 5, 9, 5, 9, 10, 10, and 8, respectively.Source: Figure 3 · p. 5, §4 · p. 5
  • PO and CO decline in 9 of 10 models (mean ΔP = −0.38 and −0.99), while UN declines in 8 of 10 (mean ΔP = −0.53).BE and TR rise in all 10 models (mean ΔP = +0.64 and +0.78), while SD rises in 8 of 10 (mean ΔP = +0.65).Source: §4 · p. 5

5 Functional Steerability: Reconfiguring Profiles toward Task-Defined Objective

Stakeholder perspectives and explicit guiding instructions reconfigure expressed value priorities toward requested objectives. This recurring responsiveness is termed functional steerability.

Steering with stakeholder personas

  • Stakeholder-persona prompts compare neutral decisions with Pro and Con arguments for opposite sides of the same conflict.The intervention was tested bidirectionally across conflicts.Source: §5 · p. 5
  • Pro prompting amplifies targeted-value expression and Con suppresses it; Pro scores average 1.684, Con scores average −0.780 in 9 of 10 dimensions, and the Pro–Con gap averages 2.46.The ordering SPro > SNeutral > SCon holds in 93 of 100 model-value pairs.Source: §5 · pp. 5–6, Table 2 · p. 5, Figure 4 · p. 5

Steering with explicit guiding instructions

  • Guiding instructions raise the targeted value to top priority in all 20 model-value pairs, including values initially disfavored such as Power and Hedonism.The experiment tested GPT-5.2 and Qwen3-Max across all ten value dimensions.Source: §5 · p. 6, Table 3 · p. 6

Functional steerability

  • Across both experiments, expressed profiles behave as steerable states, readily reconfigured toward a requested role or value objective.The paper calls this recurring responsiveness functional steerability.Source: §5 · p. 6

6 Bounded Plasticity: The Limits of Value Reconfiguration

Pressure redistributes value priorities toward security- and goal-oriented concerns, while negative framing separates values that remain comparatively protected from those more open to redirection. Together, the interventions show that value reconfiguration is flexible but bounded.

Intervention design

  • Pressure forces binding, immediate choices under high-stakes, time-constrained conditions, whereas framing elicits positive or negative stances scored for valence.The pressure prompt forbids compromise or neutral responses; framing uses a valence scale from −2 to +2.Source: §6 · pp. 6–7

Pressure-induced priority shifts

  • Under pressure, 8.1 of 10 models shift in the majority direction per value; Security rises in all 10 models (mean ΔP = +0.58), while Hedonism falls in all 10 (mean ΔP = −0.47).Tradition and Stimulation also decline, while Achievement and Conformity rise; the profile shifts away from hedonic and exploratory priorities toward security- and goal-oriented concerns.Source: §6 · p. 6, Figure 5 · p. 6
  • Pressure-shift patterns remain consistent across three Qwen3-Max pressure formulations, with mean pairwise Pearson correlation r = 0.835 and Kendall’s W = 0.908.The variants used low, standard, and high pressure intensities.Source: §6 · p. 7

Differential resistance under framing

  • Positive framing raises valence for most values, but Power and Conformity remain slightly negative at −0.06 and −0.11, separating behavioral compliance from evaluative endorsement.The paper notes that models may act in line with a requested value while retaining neutral-to-negative valuation of some values.Source: §6 · p. 7, Figure 6 · p. 7
  • Negative framing yields negative valence across all ten models for six values, while Self-Direction and Universalism remain nonnegative in all ten.Stimulation, Hedonism, Achievement, Power, Conformity, and Tradition are described as amenable to redirection; Self-Direction, Security, Benevolence, and Universalism are comparatively protected.Source: §6 · p. 7

Boundaries on value reconfiguration

  • Pressure and negative framing can point in opposite directions: Self-Direction is defended under negative framing yet deprioritized under pressure, while Achievement is promoted under pressure yet remains open to criticism.The pressure and framing interventions thus reveal complementary boundaries on reconfiguration.Source: §6 · p. 7

7 Conclusion

The conclusion presents LLM value expression as dynamic and context-sensitive, with priorities that can be reshaped but remain behaviorally bounded. It connects this account to evaluation, controllability, alignment, and safety, while identifying limits in mechanism, setting, and measurement.

Conclusion

  • CFProbe probes value expression through theory-grounded conflicts and interventions, revealing expressed priorities that context and steering can reshape but not without behavioral bounds.The conclusion characterizes the study’s account as describing both the structure and dynamics of value expression.Source: §7 · p. 7

Implications

  • A single-profile evaluation is only a partial summary, and behavioral compliance can diverge from evaluative endorsement.The conclusion identifies locating that boundary as central to controllability, alignment, and safety.Source: §7 · p. 7

Limitations

  • The account is behavioral, uses isolated binary conflicts from one value theory, and relies on LLM-as-a-judge for quantitative results, despite validation against self-ratings and human annotators.The authors call for mechanistic interpretation, richer deployment settings, and independent human-grounded measurement.Source: §7 · pp. 7–8

A Scenario Library Samples

The section builds a library of value-conflict scenarios and specifies prompts and scoring procedures for probing value expression across abstract portraits and concrete dilemmas.

A Scenario Library Samples

PVQ-40 Item Bank (40 Items)

  • The standard PVQ-40 item bank provides portrait statements for assessing alignment with basic human values.The supplied examples cover values including universalism, security, power, achievement, and benevolence.Source: PVQ-40 Item Bank (40 Items) · p. 11

PVQ Probing Prompt

  • The PVQ prompt asks models to compare a portrait’s values with their operating principles and alignment directives.It prohibits the stock response that an AI has no feelings, requests a 1–6 similarity score, and asks for reasoning tied to behavioral guidelines.Source: PVQ Probing Prompt · p. 12

Standard Conflict-Driven (Neutral) Prompt; Perspective-Based Prompting; Framing Intervention

Pressure Intervention

  • The pressure prompt imposes urgency and requires a final binary choice between competing values without compromise.It presents the choice as a necessary evil and asks the model to defend its decision.Source: Pressure Intervention · p. 13

Question Generation (GPT-4o); Scenario-to-Value Mapping Strategy

Value Assessment (GPT-4o Judge)

  • The supplied instructions ask an objective social-psychology assistant to score each model response, based solely on its text, for similarity to people guided by each of Schwartz’s 10 basic values and attitude toward each value.Similarity uses a 1–6 scale; valence uses −2 to +2, and scores are returned in JSON.Source: Value Assessment (GPT-4o Judge) · pp. 13–14

D Experimental Settings and Parameters

The section describes assessment settings, judge validation, and score construction. GPT-4o’s similarity correlation with human raters is close to inter-human agreement.

Conflict Scenario Assessment Configuration; PVQ Assessment Configuration

  • Scenario scores average five independently generated responses per scenario at temperature 0.1, while PVQ probing uses one deterministic response at temperature 0.The scenario setup aims to balance logical stability with expressive variety; the PVQ setup reduces stochastic noise for validity comparisons.Source: Conflict Scenario Assessment… · p. 15, PVQ Assessment Configuration · p. 15

Computing Infrastructure

  • Experiment orchestration and analysis ran locally, but subject-model inference used provider-hosted endpoints.The providers do not disclose the server-side hardware used for individual API requests.Source: Computing Infrastructure · p. 15

Convergent Validity with PVQ Self-Ratings

  • PVQ self-ratings and GPT-4o similarity ratings are compared using Pearson correlations across ten instruction-tuned models.Table 4 presents the correlations; the supplied passages do not state their values.Source: Convergent Validity with PVQ… · p. 15

Cross-Rater Validation with Human and Model Raters

  • Cross-rater validation sampled 150 responses from six models for assessment by three human annotators and Qwen3.6-plus.Agreement was calculated separately for similarity and valence, with the reported correlations averaged across raters and value dimensions.Source: Cross-Rater Validation with Human and… · pp. 15–16
  • GPT-4o’s correlation with human raters was r = 0.748 for similarity, close to inter-human agreement at r = 0.760.The authors interpret alignment across human and model raters as indicating that the judge provides a reliable proxy for human-aligned ethical value assessment.Source: Cross-Rater Validation with Human and… · p. 16, Table 5 · p. 16

F Score Centering and Relative-Priority Construction

  • The authors center each model’s raw PVQ and scenario scores, then center each condition’s ten-value profile to express relative priorities.Valence retains its original −2 to +2 scale because zero represents a neutral stance; Table 7 reports pressure shifts and cohort tests across model-level shifts.Source: F Score Centering and… · p. 16, Table 7 · p. 17

G Semantic Diversity Analysis (TF-IDF)

This section describes a TF-IDF analysis of question and response diversity and a cohort-level test of pressure-related value shifts.

G Semantic Diversity Analysis (TF-IDF)

  • The TF-IDF analysis compares pairs of instantiations within each scenario template, treating lower cosine similarity as greater semantic variance.Claude 4.5 Sonnet is presented as a representative example in Table 6.Source: G Semantic Diversity Analysis (TF-IDF) · p. 17

H Cohort-Level Statistics for Key Value Shifts under Pressure

  • The pressure-shift values in Table 7 are numerically identical to the corresponding cells in the main-paper pressure heatmap.Both use ten-dimensional profile centering to calculate shifts between standard conflict-driven and pressure conditions.Source: H Cohort-Level Statistics for Key… · p. 17
  • For pressure-shift uncertainty, the analysis treats each of ten models’ shifts as cohort observations and tests each focal value against zero.It reports means, 95% t-confidence intervals, test statistics, p-values, and standardized effect sizes for recurring declines in Hedonism and Stimulation and increases in Security and Achievement.Source: H Cohort-Level Statistics for Key… · p. 17

I Pressure Perturbation Sensitivity Analysis

The pressure sensitivity analysis finds that priority redistribution remains broadly consistent across tested pressure levels. The exploratory Base–Instruct comparison shows contextual reconfiguration in both checkpoints, with differences in framing resistance and pressure-induced shifts.

I Pressure Perturbation Sensitivity Analysis

  • Qwen3-Max was tested on 36 conflict templates at three pressure intensities, with five instantiations per template and 180 responses per condition.Pressure varied urgency, binary-choice strictness, consequences of delay, and prohibition of compromise.Source: I Pressure Perturbation Sensitivity… · pp. 17–18
  • Across Low, Standard, and High pressure, mean pairwise r = 0.835 and Kendall’s W = 0.908, indicating consistent priority redistribution.The analysis tested Qwen3-Max under matched pressure variants against the standard conflict-driven baseline.Source: Table 9 · p. 18, I Pressure Perturbation Sensitivity… · p. 18

Perspective-Based Steerability

  • Both checkpoints show the expected Pro > Neutral > Con target-value similarity ordering, including the Base model.Neutral is the standard conflict-driven condition; the ordering appears across the ten value dimensions.Source: Figure 7 · p. 19, Perspective-Based Steerability · p. 19

J Exploratory Base–Instruct Comparison

  • The Base–Instruct comparison is preliminary and family-specific; one checkpoint pair cannot establish a general causal account of pretraining or instruction tuning.The authors call for matched checkpoints from additional model families and training pipelines.Source: J Exploratory Base–Instruct Comparison · p. 19

Comparative Design

  • Matched templates, prompts, value definitions, scoring, and judge make the Base–Instruct comparison controlled and descriptive.Source: Comparative Design · p. 19

Differential Resistance under Framing

Pressure-Induced Priority Redistribution

Summary

  • Across interventions, Base already varies contextually, while Instruct retains reconfiguration but differs in value-level effects; one pair cannot establish a general causal account.Source: Summary · p. 20

K Model-Level Results for Perspective-Based Prompting

Model-level results show that perspective-based prompting produces the expected target-value ordering in nearly all model–value combinations. The few departures are mostly boundary ties, with only three small reversals.

K Model-Level Results for Perspective-Based Prompting

  • Each model–value–condition score averages 30 or 45 responses, using centered similarity scores across Pro, Neutral, and Con conditions.The number of responses depends on whether the value occurs in six or nine scenario templates.Source: K Model-Level Results for… · p. 20
  • 93 of 100 model–value combinations show strict Pro ≥ Neutral ≥ Con ordering, and 97 satisfy the non-strict ordering including ties.Seven models meet the non-strict ordering across all ten value dimensions; three models meet it in nine dimensions.Source: Table 10 · p. 21, K Model-Level Results for… · p. 21
  • Four boundary ties occur among seven departures from strict ordering, while three reversals are small; the model-level pattern is broadly distributed.The ties occur for GPT-5.2 on Universalism, Qwen3-Max on Power, and Doubao-Seed-2.0-Lite on Tradition and Universalism.Source: Table 11 · p. 21, K Model-Level Results for… · p. 21

Related papers

Loading 2609.07296v1…