Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 84 of the 84 references it cites were found in a registry or an index; 27 of the 37 sources its claims cite were checked for support.

Selective agreement, not sycophancy: investigating opinion dynamics in LLM interactions

Open access

Abstract

Understanding how opinions evolve is essential for addressing phenomena such as polarization, radicalization, and consensus formation. In this work, we investigate how language shapes opinion dynamics among Large Language Model (LLM) agents by simulating multi-round debates.Using our framework, we find that agent populations consistently converge toward agreement, not through sycophancy or blind conformity, but via a structured and asymmetric persuasion process. Agents are more likely to accept, and thus be persuaded by, opinions that are more agreeable relative to the discussion framing, revealing a directional bias in how opinions evolve. LLM agents selectively adopt peers’ views, showing neither bounded confidence nor indiscriminate agreement. Moreover, agents frequently produce fallacious arguments, and are significantly influenced by them: logical fallacies, especially those of relevance and credibility, play a measurable role in driving opinion change. These results not only uncover emergent behaviours in agents’ dynamics, but also highlight the dual role of LLMs as both generators and victims of flawed reasoning, raising important considerations for their deployment in socially sensitive contexts.

Overview

CoChat Summary

TL;DR

The paper introduces LODAS, an LLM-agent opinion-dynamics framework, and finds that populations of Mistral-7B and Llama-3-8B agents converge toward agreement not through sycophancy but through asymmetric persuasion: more agreeable opinions are more likely to be accepted, while more disagreeing opinions are more likely to be rejected. Agents also generate logical fallacies at around 20% of Opponent statements and are measurably influenced by them.

Key findings

  • Across 140 LLM agents, 30 iterations, and 10 runs per setting, populations consistently converged toward one or two dominant opinions rather than remaining uniform; this differed significantly from a Random null model (Welch’s t-test, p < 0.05 on most transition-matrix entries).
  • Persuasion was selective, not sycophantic: Mistral agents’ acceptance probability exceeded 60% when the Opponent held a more agreeable opinion (xO > xD) but dropped to around 20% when the Opponent held a more disagreeing view (xO < xD); Llama agents showed a more symmetric but still positively sloped P(A | Δx).
  • Initial opinion distribution had limited and transient influence: Balanced, Polarized, and Unbalanced scenarios all showed decreasing variability and convergence toward fewer opinion states, while the LLM choice was the more persistent determinant of dynamics.
  • Logical fallacies were common in Opponent statements, around 20% overall; Mistral produced slightly fewer than Llama, e.g. 15.56% under unbalanced conditions. Fallacy-induced opinion change was higher for Llama Discussants (78% same-boat, 75% different-boat) than Mistral (60% and 61%); fallacies of relevance, credibility, and logic were especially prominent.

Methods

  • LODAS is an agent-based model with N = 140 LLM agents holding discrete opinions in [0, 6] on the Ship of Theseus paradox. Agents interact pairwise at random in a mean-field setting for T = 30 iterations, with each iteration containing N interactions; one agent is Discussant and the other Opponent, and the Discussant may update by ±1.
  • The study used Mistral-7B-Instruct and Llama-3-8B, two statement framings (“the ship is the same” vs. “the ship is different”), and three initial opinion distributions (Balanced, Polarized, Unbalanced), producing 12 settings with 10 independent runs each. Agents were implemented via AutoGen AssistantAgent.
  • Analyses included opinion trajectories, entropy, standard deviation, effective number of clusters, acceptance/rejection probability matrices P(A | xD, xO) and P(R | xD, xO), signed opinion distance Δx = xO − xD, and a Random null model compared via Welch’s t-test. Logical fallacies were detected with the distilbert-base-fallacy-classification model based on Jin et al.’s 13 fallacy categories.

Limitations

  • Stated in the paper: agents are simple, lacking distinct personalities or cognitive diversity; experiments used English-aligned LLMs and may carry cultural biases; the model assumes a mean-field scenario and neglects real-world network structure; fallacious reasoning was explored only preliminarily; experiments were limited to small-scale 7–8B models due to computational constraints; broader statistical comparisons were deemed out of scope.
  • Evident from the text: the study uses only two open LLMs and a single debate topic, with no comparison to human opinion dynamics or real-world interaction data; the simulations are synthetic and the paper notes that comparing LLM-based simulations with real-world data remains future work.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1140/epjds/s13688-025-00579-1

Cites

84 of 84 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Graphical schema of LODAS. The LLM agents population is initialized as a network; each agent is an LLM instance with an initial opinion in the range [0, 6] (a). At each iteration, two agents are randomly chosen and prompted to act as Opponent and Discussant (b). The Discussant is prompted to listen to the opinion of the Opponent around the discussion statement and may then accept, reject, or ignore such opinion (c) and update their current one accordingly by ± 1 (d)
Figure 1 Graphical schema of LODAS. The LLM agents population is initialized as a network; each agent is an LLM instance with an initial opinion in the range [0, 6] (a). At each iteration, two agents are randomly chosen and prompted to act as Opponent and Discussant (b). The Discussant is prompted to listen to the opinion of the Opponent around the discussion statement and may then accept, reject, or ignore such opinion (c) and update their current one accordingly by ± 1 (d)From the PDF
Figure 2 Balanced scenario - Mistral and Llama agents opinion trends. Mistral (a)-(b) and Llama (c)-(d) opinion trends for the positive (a)-(c) and negative (b)-(d) statements. Trends are represented for Strongly Disagree (dark red), Disagree (red), Mildly Disagree (orange), Neutral (yellow), Mildly Agree (light blue), Agree (blue), and Strongly Agree (dark blue) opinions. Lines indicate the average prevalence of opinions at each time step, while the shade indicates the 95% confidence interval. Averages are computed over 10 runs
Figure 2 Balanced scenario - Mistral and Llama agents opinion trends. Mistral (a)-(b) and Llama (c)-(d) opinion trends for the positive (a)-(c) and negative (b)-(d) statements. Trends are represented for Strongly Disagree (dark red), Disagree (red), Mildly Disagree (orange), Neutral (yellow), Mildly Agree (light blue), Agree (blue), and Strongly Agree (dark blue) opinions. Lines indicate the average prevalence of opinions at each time step, while the shade indicates the 95% confidence interval. Averages are computed over 10 runsFrom the PDF
Figure 3 Balanced scenario - Mistral and Llama-agents transition matrices T. Mistral (a)-(b) and Llama (c)-(d) average (with std) transition rates from state i to state j , Tij = P ( xD ( t +1) = j | xD ( t ) = i ), for the positive (a)-(c) and negative (b)-(d) statements. Annotated cells are significant with respect to the Random Null Model ( p < 0.05) according to the t-test. Results are averaged over 10 runs
Figure 3 Balanced scenario - Mistral and Llama-agents transition matrices T. Mistral (a)-(b) and Llama (c)-(d) average (with std) transition rates from state i to state j , Tij = P ( xD ( t +1) = j | xD ( t ) = i ), for the positive (a)-(c) and negative (b)-(d) statements. Annotated cells are significant with respect to the Random Null Model ( p < 0.05) according to the t-test. Results are averaged over 10 runsFrom the PDF
Figure 4 Balanced scenario - Mistral and Llama-agents average acceptance probabilities. Mistral (orange line and matrices (a)-(b)) and Llama (blue lines and matrices (c)-(d)) average acceptance rates for the positive (solid lines and (a)-(c) matrices) and negative (dashed lines and (b)-(d) matrices) statements. Top panel: acceptance probability P ( A | Δ x ) as a function of opinion distance Δ x = xO -xD . Bottom panel: P ( A | oD , oO ) averaged over 10 runs
Figure 4 Balanced scenario - Mistral and Llama-agents average acceptance probabilities. Mistral (orange line and matrices (a)-(b)) and Llama (blue lines and matrices (c)-(d)) average acceptance rates for the positive (solid lines and (a)-(c) matrices) and negative (dashed lines and (b)-(d) matrices) statements. Top panel: acceptance probability P ( A | Δ x ) as a function of opinion distance Δ x = xO -xD . Bottom panel: P ( A | oD , oO ) averaged over 10 runsFrom the PDF
Figure 5 Average Effective Number of Clusters C . Average effective number of clusters for Llama (blue) and Mistral (orange), and for the Same (solid) and Different (dashed) statements. Results are averaged over 10 runs, the shaded area represents the standard deviation
Figure 5 Average Effective Number of Clusters C . Average effective number of clusters for Llama (blue) and Mistral (orange), and for the Same (solid) and Different (dashed) statements. Results are averaged over 10 runs, the shaded area represents the standard deviationFrom the PDF
Figure 6 Acceptance probability P ( A | Δ x ) as a function of opinion distance Δ x = xO -xD. Each marker represents the average probability that a Discussant agent accepts-i.e., moves toward-the Opponent ' s opinion, as a function of the opinion distance Δ x = xO -xD . Lines correspond to different initial opinion distributions: Balanced (blue), Polarized (orange), and Unbalanced (green). Rows distinguish between the two language models: Mistral (top) and Llama (bottom). Columns refer to the direction of the statement: Same (left) and Different (right). Results are averaged over 10 independent runs; shaded areas indicate standard deviations across runs
Figure 6 Acceptance probability P ( A | Δ x ) as a function of opinion distance Δ x = xO -xD. Each marker represents the average probability that a Discussant agent accepts-i.e., moves toward-the Opponent ' s opinion, as a function of the opinion distance Δ x = xO -xD . Lines correspond to different initial opinion distributions: Balanced (blue), Polarized (orange), and Unbalanced (green). Rows distinguish between the two language models: Mistral (top) and Llama (bottom). Columns refer to the direction of the statement: Same (left) and Different (right). Results are averaged over 10 independent runs; shaded areas indicate standard deviations across runsFrom the PDF
Figure 7 Average logical fallacies proportions for different experiment configurations across multiple runs. Figures (a)-(b) refer to Llama, while (c)-(d) refer to Mistral. (a)-(c) refer to Llama (a) and Mistral (c) agents discussing the same boat statement, (b)-(d) refer to the different boat statement. Error bars represent standard deviation across 9 runs. The x-axis uses the following abbreviations: R (fallacy of relevance), L (fallacy of logic), CR (circular reasoning), C (fallacy of credibility), G (faulty generalization), I (intentional), and FC (false causality)
Figure 7 Average logical fallacies proportions for different experiment configurations across multiple runs. Figures (a)-(b) refer to Llama, while (c)-(d) refer to Mistral. (a)-(c) refer to Llama (a) and Mistral (c) agents discussing the same boat statement, (b)-(d) refer to the different boat statement. Error bars represent standard deviation across 9 runs. The x-axis uses the following abbreviations: R (fallacy of relevance), L (fallacy of logic), CR (circular reasoning), C (fallacy of credibility), G (faulty generalization), I (intentional), and FC (false causality)From the PDF

Verification

Claims the paper makes

16 Valid · 5 Not valid · 7 Not checked · 84 Own results and interpretations

54 more claims

Create a free account to see all 112 claims and to verify the 4 references still open.

See all claims — free account
  • Probably not a substantial claim. 19
  • It may state two separate claims as one. 43
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 47 references

Research works printed but not linked: 2

  1. Likert R (1932) A technique for the measurement of attitudes. Arch Psychol 22 140:55
  2. Copi IM, Cohen C, MacMahon K (2013) Introduction to logic, 14 edn. Pearson custom library. Pearson Education, Upper Saddle River

Cited by

See all 9

Metadata

Authors
Cau, E., Pansanella, V., Pedreschi, D., & Rossetti, G.
Affiliations
  • University of Pisa (IT)
  • Istituto di Scienza e Tecnologie dell'Informazione "Alessandro Faedo" (IT)
Corresponding author
Cau, Erica
Venue
EPJ Data Science
Published
2025-08-15
Volume
14
Issue
1
ISSN
2193-1127 (ISSN-L)
OpenAlex
W4413415991
Cited by
9
References
47
Citations by year
2026: 7 · 2025: 2
FWCI
3.66
Type
journal article
Language
en
Open access
Licence
Locations
  • EPJ Data Science (publishedVersion, cc-by-nc-nd)
  • CINECA IRIS Institutial research information system (University of Pisa) (submittedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • European Commission (IR0000013, 871042)
  • NextGenerationEU
Topics
  • Physical Sciences › Physics and Astronomy › Statistical and Nonlinear Physics › Opinion Dynamics and Social Influence
  • Physical Sciences › Computer Science › Artificial Intelligence › Multi-Agent Systems and Negotiation
  • Physical Sciences › Computer Science › Artificial Intelligence › Sentiment Analysis and Opinion Mining
Keywords
opinion dynamics, LLM agents, persuasion, sycophancy, conformity, logical fallacies, opinion change, polarization, radicalization, consensus formation, emergent behavior

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Selective agreement, not sycophancy: investigating opinion dynamics in LLM interactions. https://app.cochat.ai/papers/10.1140/epjds/s13688-025-00579-1
The paper
Cau, E., Pansanella, V., Pedreschi, D., & Rossetti, G. (2025). Selective agreement, not sycophancy: investigating opinion dynamics in LLM interactions. EPJ Data Science. https://doi.org/10.1140/epjds/s13688-025-00579-1

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.