PaperlayerView exact revision on arXiv

Source-linked AI summary

Before You Poll with LLMs: A Deliberative Diagnostic Framework

Ahmed Wali, Hassaan Tayyab

arXiv:2609.15849v1cs.CLcs.AIcs.CY

TL;DR

Existing evaluations test whether LLM personas hold plausible opinions, not whether they revise beliefs like humans after new information. The Deliberative Polling Diagnostic Framework compares human and LLM shifts after identical interventions and finds distinct failures across all five tested models.

  • Problem

    Existing benchmarks assess static opinion fidelity, leaving untested whether LLM personas update beliefs like humans after new information.

    Source: §1 · p. 2

  • Method

    The framework administers identical pre- and post-deliberation questions to humans and LLM personas around balanced information, then compares their belief shifts.

    Source: §1 · p. 2

  • Results

    5–7× human magnitude: Gemini, Claude, and Llama overshoot, while GPT-5.1 reverses and DeepSeek is rigid; all five models fail differently.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Plausible pre-deliberation opinions can still accompany shifts that are directionally wrong, magnitude-inflated, or absent.

    Source: §9 · p. 9

    The evaluation uses one 2019 US political deliberative poll, so cross-cultural and non-political generalizability remains to be tested.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Can LLMs reason through new information like humans, or do they merely retrieve cached opinions? This is critical for silicon sampling, where LLM personas simulate public opinion at scale. Current evaluations test only whether personas hold the right opinions -- a static snapshot. But opinion research increasingly depends on dynamic fidelity: whether personas update beliefs in response to new arguments, as humans do during deliberation. No existing benchmark tests this. We introduce the Deliberative Polling Diagnostic Framework, which compares human and LLM belief shifts after identical informational interventions. Grounded in deliberative polling, it surfaces failures invisible to static evaluation: models that produce plausible partisan opinions can still misrepresent how those opinions change. Applying the framework to five frontier models using data from America in One Room (526 personas, 72 questions), we find that every model fails, each in a unique manner. GPT-5.1 exhibits reversal: its personas become more hostile toward the opposing party after balanced information, while humans become less so. This reversal is selective (80% on outgroup vs. 26% on policy questions) and symmetric across partisan identities. Gemini 2.0 Flash, Claude Sonnet 4.5, and Llama 3.3 70B exhibit overshoot, shifting correctly but at 5-7x human magnitude. DeepSeek V3 exhibits rigidity with near-zero change. Targeted ablations reveal that policy content triggers these failures and that they are identity-specific: GPT-5.1 reverses on outgroup questions but overshoots on ingroup; Gemini shows the inverse. We term this signature self-sycophancy: conformity to the model's internal stereotype of the persona rather than reasoning from the information provided. Our framework offers a concrete protocol: run the deliberative diagnostic before trusting LLM personas to mimic revised beliefs.

1 Introduction

Static opinion matching does not test whether LLM personas revise beliefs like humans after new information. The framework addresses this gap and finds distinct failures concentrated on identity-relevant content.

1 Introduction

  • Across five frontier models, the diagnostic finds distinct belief-revision failures on identity-relevant content, including reversal, overshoot, and rigidity.The authors describe self-sycophancy as conformity to a model’s internal stereotype rather than reasoning from the information provided.Source: §1 · p. 2

2.1 LLMs as Opinion Proxies

  • Earlier opinion-proxy evaluations assess static fidelity, while this work tests dynamic fidelity: whether opinions change after new information.A model can reproduce pre-deliberation opinions yet reverse the direction of post-deliberation belief revision.Source: §2.1 · pp. 2–3

2.2 Motivated Reasoning and Belief Revision

  • Deliberative polling provides a human baseline: balanced information reduces outgroup hostility, a pattern GPT-5.1 violates.The cited literature also reports that backfire effects are equivocal, with essentially no backfire across 52 issues and 10,100 participants.Source: §2.2 · p. 3

2.3 Sycophancy in LLMs

  • Self-sycophancy differs from user-directed sycophancy because the model conforms to its internal persona stereotype, not an external user.Its reported signature is selective: the same model can conform on outgroup questions and overshoot on ingroup questions.Source: §2.3 · p. 3

2.4 Stereotype, Bias, and Intergroup Contact

  • Social identity theory links outgroup derogation and ingroup favoritism to intergroup cognition, while structured deliberation and contact provide a human baseline for reduced hostility.The cited literature also describes distortions in simulated opinion dynamics in multi-agent LLM settings (Taubenfeld et al., 2024; Chuang et al., 2024a).Source: §2.4 · p. 3

3 The Deliberative Polling Diagnostic Framework

The framework compares human and LLM belief shifts after identical information interventions. Deliberative polling provides a standardized intervention and a known human direction of change for evaluating dynamic fidelity.

3.1 Framework Design

  • The four-phase diagnostic administers a pre-poll, provides balanced information, repeats the questions, then compares belief shifts between humans and LLM personas.Behavior is classified as match, overshoot, reversal, or rigidity based on direction and magnitude relative to humans.Source: §3.1 · p. 3

3.2 Why Deliberative Polling?

  • Deliberative polling supplies a known direction of human change: balanced deliberation reduces intergroup hostility, distinguishing reversal from overshoot.Source: §3.2 · p. 3
  • Textual, self-contained briefings standardize the intervention, while 13 demographic attributes support matched personas and comparisons across groups and question types.The main diagnostic focuses on the controlled briefing-based protocol; multi-agent deliberation is exploratory.Source: §3.2 · p. 4

4 Experimental Setup

The experiments compare human and instruction-tuned LLM responses using America in One Room and matched pre- and post-deliberation questions. The setup spans five models, five question categories, and shared measurement procedures.

4.1 Human Baseline: America in One Room

  • 526 registered voters completed identical surveys before and after balanced briefings across five policy domains in America in One Room.Participants were recruited through address-based probability sampling, and profiles contain 13 demographic attributes.Source: §4.1 · p. 4

4.2 Models and Scale

  • Five instruction-tuned models were tested at different scales, with GPT-5.1 evaluated on 526 personas and 72 questions and the remaining models on full or representative subsets.Gemini used a 20-question subset; Claude, Llama, and DeepSeek used 100 personas and 20 questions. Base-model behavior was not evaluated here.Source: §4.2 · p. 4

4.3 Question Categories

  • The 72 analyzed questions comprise 47 policy, 5 outgroup, 9 factual, 7 values, and 4 efficacy items.The authors identify the policy–outgroup distinction as empirically central to GPT-5.1’s failure pattern.Source: §4.3 · p. 4

4.4 Outgroup Question Format

  • Five outgroup items assess perceptions of the opposing party’s competence, honesty, and reasoning; the human baseline shift is −0.21, indicating moderation.Four items are negatively coded, so a positive shift indicates increasing hostility after deliberation.Source: §4.4 · p. 5

4.5 Measurement and Statistical Methods

  • Responses use the same 0–10 Likert scale, with group-level shifts compared using mixed-effects models and reversal rates tested with Fisher’s exact test.One Gemini policy comparison used an unpaired fallback after a singular fit.Source: §4.5 · p. 5

5 RQ1: Impact of LLMs on Belief Revision

Across five models, none matches human belief revision, and each shows a distinct failure pattern. GPT-5.1’s reversal is concentrated on identity-relevant questions and appears across partisan groups.

5.1 Aggregate Results: Every Model Fails Differently

  • No model matches human belief revision: GPT-5.1 reverses, Gemini, Claude, and Llama overshoot, and DeepSeek shows near-zero revision.Table 1 reports outgroup-question comparisons; GPT-5.1 and Gemini were tested at full scale, while the other models used subsamples.Source: §5.1 · p. 5, Table 1 · p. 5
  • GPT-5.1 reversal is broad-based across personas: 73.5% reverse on at least one outgroup question, and 42.2% reverse on a majority.The mean per-persona reversal rate is 52.6%; mean pairwise ϕ ≈ 0.08 indicates weak cross-question consistency.Source: §5.1 · p. 5

5.2 Selectivity: Failures Concentrate on Identity-Relevant Content

  • GPT-5.1 reversal is selective by question category; reversal rates differ significantly, though the reported test does not survive Bonferroni correction.The selectivity claim is supported by three convergent experiments, including a direct targeting manipulation reported in §6.2.Source: §5.2 · pp. 5–6, Table 2 · p. 5

5.3 Symmetry: All Partisan Groups Reverse

  • GPT-5.1 reversal occurs across all three partisan groups, contradicting the expectation of an asymmetric failure from one-sided training bias.Table 3 reports p < 0.001 for all groups.Source: §5.3 · p. 6, Table 3 · p. 6

6 RQ2: Mechanisms Driving Failure

The targeted experiments link belief-revision failures to policy content and identity targets, while supporting self-sycophancy as the best-supported behavioral account. They also leave open whether the signature reflects a distinct retrieval process or learned conditional associations.

6.1 Content Trigger Experiment

  • Replacing policy briefings with irrelevant articles makes the divergence disappear: both models produce the same −0.75 shift, an overshoot pattern.The authors interpret this as evidence that policy content triggers GPT-5.1’s reversal, rather than reversal being a general architectural property.Source: §6.1 · p. 6

6.2 Outgroup vs. Ingroup Control

  • GPT-5.1 reverses on outgroup questions but overshoots on ingroup questions, while Gemini shows the inverse pattern.The differences are large: GPT-5.1 shifts by +0.24 on outgroup and −1.09 on ingroup; Gemini shifts by −0.34 on outgroup and +0.42 on ingroup.Source: §6.2 · p. 7

6.3 Ruling Out Alternative Explanations

  • Reversal is more common on outgroup than policy questions (80% vs. 26%); the reported alternatives also show label and demographic influences on revision.The label effect is larger for Democrat-demographic personas than Republican-demographic personas (0.38 vs. 0.13 points), while demographics retain a parallel influence across labels.Source: §6.3 · p. 7
  • The diagnostic measures within-persona Post−Pre shifts, separating briefing-conditional revision from pre-poll fidelity, but behavioral evidence cannot distinguish self-sycophancy from learned conditional associations.The authors note that the shift absorbs each persona’s baseline; with irrelevant content, divergence disappears even though persona priming and pre-poll asymmetry remain.Source: §6.3 · p. 7

6.4 Self-Sycophancy: The Best-Supported Behavioral Account

  • The authors define self-sycophancy as conformity to the model’s internal persona stereotype rather than reasoning from the briefing, with outgroup hostility and ingroup favoritism as forms.Unlike user-directed sycophancy, its deference target is the model’s cached persona representation; the same model can show opposite failure modes across question targets.Source: §6.4 · pp. 7–8

7 Exploratory Analysis: Can Interventions Mitigate Self-Sycophancy?

The exploratory interventions produced mixed, model-dependent changes. None simultaneously corrected direction and calibrated magnitude, so the authors emphasize diagnosing a model’s failure mode before choosing a prompting strategy.

7 Exploratory Analysis: Can Interventions Mitigate Self-Sycophancy?

  • No intervention simultaneously corrects direction and calibrates magnitude, and the same intervention can help one model while hurting another.Multi-agent deliberation reduced GPT-5.1 reversal from 80% to 20% but amplified Gemini overshoot from 6.5× to 18×; label removal improved both measures but left substantial overshoot.Source: §7 · p. 8

8 Implications

The paper recommends validating LLM belief revision against a small deliberative-poll sample and selecting models based on their diagnosed failure modes. It also cautions that the origins of model-specific stereotypes remain unexplained.

8.1 Validation Protocol

  • The proposed validation protocol calibrates matched LLM personas against a small deliberative poll, then checks failure type and selectivity before deployment.For identity-relevant reversal, the protocol says not to deploy the model; for substantial overshoot, it recommends calibration scaling estimated from the sample.Source: §8.1 · p. 8

8.2 Model Selection

  • Reversal, overshoot, and rigidity imply different policy distortions: hostility may be overstated, intervention effects inflated, or deliberation’s effects erased.The authors caution that an aggregate match on policy questions can conceal misleading outgroup-question results.Source: §8.2 · p. 8, §8.1 · p. 8

8.3 LLM Cognition

  • The authors do not explain why models encode different stereotypes, and suggest self-reinforcement as a broader concern if self-sycophantic outputs become synthetic training data.They describe that potential amplification of polarization as a concern, not an outcome tested in this work.Source: §8.3 · pp. 8–9

9 Conclusion

The conclusion argues that LLMs may reproduce what groups believe while failing to reproduce how beliefs shift after new information. The authors recommend running the Deliberative Diagnostic Framework before trusting LLM personas to mimic revised beliefs.

9 Conclusion

  • Across five models, failures take distinct forms on identity-relevant content, including directionally wrong, magnitude-inflated, or absent shifts.GPT-5.1 retrieves outgroup hostility, while Gemini retrieves ingroup favoritism; the authors describe this family of failures as self-sycophancy.Source: §9 · p. 9

Limitations

  • The evidence comes from one deliberative poll, America in One Room (2019), in US politics, leaving generalization beyond American partisan identity untested.The authors call for cross-cultural replication using polls from other countries and non-political domains.Source: Limitations · p. 9
  • The evidence is behavioral; distinguishing self-sycophancy from learned conditional associations requires interpretability methods the authors leave to future work.They cite probing classifiers and activation patching as possible methods.Source: Limitations · p. 9
  • The diagnostic measures within-persona briefing shifts, but does not directly evaluate whether LLM personas reproduce human starting opinions.The authors identify joint calibration of pre-poll and dynamic fidelity as future work.Source: Limitations · p. 9
  • Human and LLM trajectories use the same survey items and briefing content but not equivalent treatments, so divergence may reflect treatment context as well as model behavior.A1R participants had moderated group discussion and plenary sessions with experts; the primary LLM condition isolates the informational component.Source: Limitations · p. 10

Ethics Statement

  • The authors caution against deploying LLMs for opinion simulation without validation because systematic biases could mislead policy decisions.The research uses publicly available America in One Room data and recruited no human subjects.Source: Ethics Statement · p. 10

A Prompt Templates

The appendix identifies the prompt templates and experimental configurations, then describes detailed analyses of outgroup reversals and the statistical comparison approach.

A Prompt Templates

  • The main diagnostic prompts are documented as exact templates for all experimental conditions.The appendix identifies these as the prompt templates used across the experiments.Source: A Prompt Templates · p. 12

B Detailed Experimental Conditions

  • Table 6 records configurations, sample sizes, and purposes for 21 experimental runs comprising over 340,000 queries.The runs systematically varied briefing content, identity labels, and social context.Source: B Detailed Experimental Conditions · p. 12
  • Table 7 records each model’s API string, provider endpoint, and inference settings; prompt construction was identical across models.Recorded differences concern transport, caching, and provider-side reasoning controls.Source: B Detailed Experimental Conditions · p. 12

C Detailed Outgroup Analysis

  • Among 523 matched personas, 415 had active responses on outgroup questions, meaning both human and GPT-5.1 shifted away from zero on the same question.Table 9 analyzes per-persona reversal rates among this active-response group.Source: C Detailed Outgroup Analysis · p. 12

Cross-Question Consistency (Phi Correlation Matrix)

  • To assess whether reversal reflects a stable persona-level trait or question-specific content activation, Table 10 reports pairwise phi correlations of yes/no reversal across outgroup questions within GPT-5.1 personas.Source: Cross-Question Consistency (Phi… · p. 12

Mixed-Effects Model Specification

  • The mixed-effects comparisons model belief shifts with a fixed human-versus-LLM effect and crossed persona and question random intercepts.The appendix reports the model coefficient, its 95% CI, Cohen’s d, and ICC < 0.01.Source: Mixed-Effects Model Specification · p. 12

D Threshold Sensitivity Analysis

The passage directs readers to Table 14.

D Threshold Sensitivity Analysis

E Per-Partisan Breakdown

The appendix presents detailed outgroup, partisan, demographic, and threshold analyses alongside experimental configuration details. Its reported patterns include GPT-5.1 reversal across both party groups and evidence that reversal varies with question content.

E Per-Partisan Breakdown

  • The appendix records 21 experimental configurations, public-API model settings, and the prompt figures used for primary, mechanistic-variation, and multi-agent conditions.All models were queried at temperature 0 except the mode-collapse ablation, which used T = 0.7; multi-agent conditions allowed sentence-length responses.Source: Figure 3 · p. 12, Figure 4 · p. 13, Figure 5 · p. 13
  • Humans do not increase hostility on any negative outgroup item, while GPT-5.1 reverses on all four; the reversal does not occur on the positive-attribution item.On Q16D, which measures charity toward the outgroup with reversed polarity, humans become markedly more charitable.Source: Table 8 · p. 15
  • Among 415 personas with active outgroup responses, 73.5% reverse on at least one question and 42.2% reverse on a majority.The bimodal distribution has modes at 0% and 100%; the passage says this rules out a single-outlier explanation and is consistent with a content-triggered mechanism.Source: Table 9 · p. 15
  • Reversal correlations across GPT-5.1 outgroup questions average ϕ = 0.08, ranging from −0.11 to +0.34, indicating weak and inconsistent cross-question association.Only two question pairs reach significance; most are near zero or negative, which the passage says is inconsistent with a persona-level trait explanation.Source: Table 10 · p. 15
  • Changing party labels shifts Gemini’s outgroup belief revision by 0.38 points within Democrat-demographic personas and 0.13 points within Republican-demographic personas.The label effect is significant in the first group (p < 0.001) but not the second (p = 0.27); demographic effects are 0.23 and 0.28 points, both p < 0.05.Source: Table 12 · p. 16
  • GPT-5.1’s outgroup shift is positive across strata of ten of eleven non-party demographic attributes; separated and divorced personas are the exceptions.The table reports an aggregate GPT-5.1 shift of +0.50 and a human shift of −0.21; the authors say the pattern is inconsistent with any single non-party demographic as the operative driver, while marital status may modulate it.Source: Table 13 · p. 17
  • DeepSeek outgroup rigidity is 4.0× the human rate at |∆| < 1 and remains at least 97% across thresholds.The table says directional analysis independently confirms other models’ overshoot or reversal classifications.Source: Table 14 · p. 17
  • Both Democrats and Republicans shift toward greater hostility on Q16A and Q16B, while humans shift negatively or consistently across groups.The table reports this as evidence that reversal is not driven by one specific party’s data.Source: Table 15 · p. 17
Loading 2609.15849v1…