Before You Poll with LLMs: A Deliberative Diagnostic Framework

AI summary

Overview

Research area: Natural Language Processing / LLM evaluation, with ties to computational social science and political opinion research.

Technical level: Intermediate. The conceptual argument is accessible to anyone familiar with LLM benchmarking, but the paper relies on mixed-effects modeling, effect-size reporting, and deliberative polling methodology that reward some statistical background.

Scope: The paper introduces and applies a diagnostic framework that tests whether LLM personas revise their beliefs in the same direction and magnitude as humans after receiving identical balanced information, evaluating five frontier models against data from the America in One Room deliberative poll.

What This Paper Is About

Researchers increasingly use LLM personas as stand-ins for survey respondents, a practice called silicon sampling, but existing evaluations only check whether a persona holds the right opinion at a single moment. They never check whether that persona updates its opinion the way a human would when given new arguments. This paper builds a framework that administers the same pre-poll, balanced briefing, and post-poll sequence to both humans and LLM personas, then compares how much and in what direction each side's beliefs shift.

Key Contributions

1. A diagnostic framework for dynamic fidelity. The Deliberative Polling Diagnostic Framework repurposes deliberative polling methodology into an evaluation protocol with four phases (pre-poll, briefing, post-poll, comparison) and a four-way behavior taxonomy: match, overshoot, reversal, and rigidity.

2. Documentation that all five frontier models fail, each in a distinct way. Using 526 personas and 72 questions from America in One Room, the authors show GPT-5.1 reverses, Gemini/Claude/Llama overshoot, and DeepSeek stays rigid. No model reproduces human belief revision.

3. Identification of self-sycophancy as a behavioral failure mode. Through content-trigger, ingroup/outgroup, and label-swap ablations, the authors characterize a failure distinct from user-directed sycophancy: the model defers to its own internal stereotype of the persona rather than reasoning from the briefing material.

4. A concrete practitioner protocol. A four-step validation procedure any organization can run before deploying LLM personas for opinion simulation, including a specific warning about selectivity — models can pass aggregate metrics while failing on exactly the identity-relevant questions that matter most.

Main Findings

• Every model fails differently. On outgroup questions, the human baseline shift is −0.21 (moderation). GPT-5.1 shifts +0.50 (reversal — becoming more hostile after balanced information). Gemini 2.0 (−1.36), Claude Sonnet 4.5 (−1.04), and Llama 3.3 70B (−1.42) shift in the right direction but at 5–7× human magnitude (overshoot). DeepSeek V3 shifts +0.02 (rigidity). All model–human comparisons survive Bonferroni correction except DeepSeek (p = 0.051). The diversity of failure modes points toward training-specific factors rather than a single architectural cause.

• Failures are selective to identity-relevant content. GPT-5.1 reverses on 80% of outgroup questions (4/5) but only 26% of policy questions (12/47). The authors interpret this as a declarative–procedural gap: policy questions can be answered by retrieving stored positions, while outgroup questions require simulating how new information changes attitudes.

• Reversal is symmetric across partisan identities. Democrat, Republican, and Independent personas all reverse (all p < 0.001). One-sided training bias would predict asymmetric failure; the symmetry instead suggests a uniform implicit theory applied regardless of the persona's party.

• Reversal is broad-based, not subgroup-driven. 73.5% of personas reverse on at least one outgroup question, and 42.2% reverse on a majority. Mean per-persona reversal rate is 52.6%, though cross-question consistency is weak (mean pairwise φ ≈ 0.08).

• Policy content is the trigger. Replacing policy briefings with irrelevant Wikipedia articles (pasta, chess, gardening) eliminates the divergence entirely — both GPT-5.1 and Gemini produce identical overshoot (Δ = −0.75). The model can moderate; it simply does not when presented with politically relevant material.

• The same model produces opposite failures depending on identity target. With explicit party labeling, GPT-5.1 reverses on outgroup questions (Δ = +0.24) but overshoots on ingroup (Δ = −1.09), t = 12.29, d = 0.43. Gemini shows the exact inverse: near-human on outgroup (Δ = −0.34) but reversing on ingroup (Δ = +0.42), t = −12.60, d = −0.44. Two models, two different stereotypes, one underlying mechanism.

• Six alternative explanations were tested and ruled out, including briefing format, mode collapse (temperature 0.7 preserves the failure, paired r = 0.77), partisan training bias, persona construction artifacts, demographic confounds, and a label-swap experiment showing party label is a strong retrieval cue specifically for Democrat-demographic personas (0.38-point shift vs. 0.13 for Republican-demographic personas).

• No intervention fixes both direction and magnitude. Multi-agent deliberation reduces GPT-5.1's reversal rate from 80% to 20% but amplifies Gemini's overshoot from 6.5× to 18×. Chain-of-thought and persona-first ordering eliminate Gemini's reversal but leave overshoot intact at roughly 5.7×. Label removal offers the most balanced improvement but the model still substantially overshoots.

Methodology in Plain English

The researchers took an existing experiment on real people and ran a mirror version with AI. In America in One Room, 526 registered voters answered surveys before and after reading balanced briefing materials on five policy areas. The authors rebuilt each of those 526 voters as an LLM persona using 13 demographic attributes (party, age, race, education, income, and so on), then gave the personas the same questions and the same briefings under the same conditions.

Crucially, the LLM was never shown its own pre-briefing answer when answering the post-briefing question — this matches the human protocol and ensures both sides are measured identically. The key quantity is the shift Δ = Post − Pre, computed separately for humans and for personas.

Behavior on each question falls into one of four categories: match (same direction and comparable size), overshoot (right direction, inflated size), reversal (wrong direction), or rigidity (essentially no movement).

To figure out why the failures happen, the authors ran targeted ablations. They swapped policy briefings for Wikipedia articles on unrelated topics to test whether political content triggers the failure. They rewrote questions to explicitly name the target party (Democrats vs. Republicans) rather than using vague phrasing like "those people," to test whether the model behaves differently toward ingroups versus outgroups. They stripped the party label from the persona description to see how much the label alone drives behavior.

Scale: five frontier models, 21 experimental conditions, over 340,000 queries. GPT-5.1 ran at full scale (526 personas × 72 questions); Gemini ran at full scale on a 20-question subset; the others ran on 100-persona subsamples. All queries used temperature 0 with integer responses on a 0–10 scale.

Why This Matters

Impact on research. Every existing benchmark for LLM opinion simulation tests static fidelity — whether a persona reproduces human opinions at one point in time. This paper demonstrates that a model can pass those benchmarks cleanly while producing directionally wrong belief updates. It opens a second evaluation axis (dynamic fidelity) that the silicon sampling literature has not previously measured.

Real-world applications:

• Government policy consultation. An agency considering whether to convene citizen panels might query 10,000 LLM personas instead. GPT-5.1 would tell them deliberation increases hostility, leading them to conclude public consultation backfires — a conclusion that is a model artifact, not a finding.

• Survey research and polling. Pollsters using LLM personas to pre-test question wording or estimate opinion shifts would get effect sizes inflated 5–7× by Gemini, Claude, or Llama, badly overstating how persuadable the public is.

• Public health and science communication. The same overshoot pattern would mislead campaigns estimating how much a messaging intervention moves attitudes on vaccines, climate, or health behaviors.

• Commercial market research. Brand and product researchers substituting LLM personas for focus groups inherit the same selectivity problem, especially on any question touching group identity.

Industry relevance. The multi-agent intervention result is particularly pointed: the same intervention helps one model and hurts another. A practitioner cannot choose a prompting strategy without first diagnosing which failure mode their specific model exhibits, which makes the diagnostic framework a prerequisite rather than a nice-to-have. The paper also raises a self-reinforcement concern — if self-sycophantic outputs get recycled into training data, the encoded stereotypes could amplify polarization beyond its prevalence in human populations.

Future Directions

1. Cross-cultural and cross-domain replication. The current results rest on a single 2019 US political deliberative poll. Testing deliberative polls from other countries and non-political domains (health, consumer behavior) would establish whether self-sycophancy generalizes beyond American partisan identity.

2. Mechanistic interpretability of the failure. The evidence here is purely behavioral and cannot distinguish a structurally distinct stereotype-retrieval process from learned conditional associations in training data that produce the same signature. Probing classifiers and activation patching would adjudicate between these.

3. Base-model testing to locate the origin. All five tested models are instruction-tuned. Concurrent work shows base models exhibit parallel failures, but confirming whether self-sycophancy exists pre-RLHF would clarify whether it originates in pretraining data or in alignment procedures.

4. Joint calibration of pre-poll and dynamic fidelity. The diagnostic deliberately measures within-persona shifts to absorb baseline differences, which means pre-poll fidelity remains a separate, unaddressed axis. Belief-network conditioning and related methods for calibrating both simultaneously are an open problem. A repeated-prompting baseline at temperature 0 would also provide a proper floor against which to scale observed shifts.

Target Audience

Primary: NLP and LLM evaluation researchers, particularly those working on persona simulation, sycophancy, or alignment. The framework and behavior taxonomy are directly reusable.

Secondary: Computational social scientists and survey methodologists who use or are considering LLM personas as synthetic respondents. The validation protocol in Section 8.1 is written for exactly this audience.

Also relevant to: AI governance and policy analysts assessing whether LLMs are suitable for democratic consultation tools; industry practitioners in market research or public opinion who need to know the specific failure signature of the model they are deploying.

Less suited for: Readers looking for a new model architecture or training technique — this is an evaluation and diagnostic contribution, not a methods paper for building better LLMs.

Authors’ abstract

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.