Bias Amplification in Multi-Agent Network: How Biased Agents Shape Opinions and Rhetoric
Omran Berjawi, Giuseppe Fenza, Rida Khatoun
cs.LG
Quick verdict
Biased LLM minority (≤20%) shifts neutral agents beyond an FJ baseline with rhetoric decoupled from numbers; plausible, but single-model and one-sided.
Summary
The paper simulates opinion dynamics among 50 fully connected LLM agents (Llama 3.2-8B, LangChain, temperature 0.7) on four contested topics, seeding a biased minority fixed at the extreme +2 stance (0-20% of the population) while neutral agents, initialized at {-1, 0, 1}, update one category step per round via structured prompts exposing every agent's stance and justification. Over 10 rounds and 15 runs per configuration, neutral agents' mean opinion shift grows monotonically with bias share, reaching 1.49-1.92 at 20% bias versus at most 0.47 for a matched Friedkin-Johnsen baseline (biased agents fully stubborn, lambda_n=0.5); most movement occurs within three rounds. SBERT cosine alignment of neutral agents' justifications with biased agents' text also rises monotonically with exposure and, the authors argue, is partially decoupled from numeric convergence - rhetoric homogenizes even where opinion shifts are moderate. Notably, the vaccine topic drifts +0.66 with zero biased agents, revealing an intrinsic model prior. Limitations include a single model and framework, one-sided (+2 only) anchors conflating minority influence with model bias, prompts exposing numeric anchor values, alignment measured only at the final round, fully connected topology only, no significance testing, and no released code. The authors position the work as a proof-of-concept.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Even a small minority (2-10 of 50) of agents persistently fixed at +2 shifts neutral LLM agents' opinions monotonically with bias share, with most change occurring within the first three rounds.
- Under matched configurations, Llama 3.2-mediated interaction produces faster and larger convergence toward the biased anchor than the FJ baseline (e.g., mean shift 1.49 vs 0.47 at Ab=10 on AI safety regulation).
- Semantic alignment of neutral agents' textual justifications with biased agents' text (SBERT cosine) increases monotonically with biased exposure across all four topics.
- Rhetorical convergence is partially decoupled from numeric convergence: agents adopt the biased minority's vocabulary even when their numerical opinion shifts remain moderate.
- Llama 3.2 exhibits an intrinsic pro-agree prior on some topics, drifting +0.66 on vaccine mandates with zero biased agents present.
Strengths
Controlled causal design isolating minority size
The biased share is varied systematically (Ab in {0,2,4,6,8,10} of N=50) with population, network, topics, and initialization held constant, an Ab=0 baseline anchors comparisons (Section 3.2), and the FJ comparison matches stubbornness explicitly (lambda_b=1, Section 4.4) so the LLM-vs-classical contrast is attributable to the update mechanism.
Novel dual-layer measurement of opinions and rhetoric
Jointly tracking numeric opinions and SBERT cosine alignment of justifications (Section 4.3) yields the paper's most interesting observation: alignment rises substantially at Ab=4-6 in topics where numeric shifts remain moderate (Section 5.2), a decoupling not previously documented in LLM opinion-dynamics work.
Replication effort and reporting transparency
Candid, specific limitations section
Weaknesses
One-sided anchors confounded with intrinsic model prior
Alignment metric lacks a control baseline
Prompts mechanically expose numeric anchors
FJ baseline comparison is parameter-sensitive and under-specified
Thin statistical inference for the central claims
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What happens when biased LLM agents are added to a group of neutral LLM agents?
In this simulation, a minority of agents fixed at an extreme stance steadily pulls the neutral majority's opinions toward that stance. With 10 biased agents among 50 (20%), neutral agents' opinions moved roughly 1.5-1.9 points on a -2 to +2 scale, and most of the movement happened within the first three discussion rounds.
Do LLM agents converge faster than classical opinion dynamics models?
According to this paper's comparison, yes. Under identical networks and bias levels, Llama 3.2-driven discussion shifted opinions much further toward the biased position (e.g., a mean shift of 1.49 versus 0.47 at 20% bias) than the classical Friedkin-Johnsen model with equally stubborn biased agents.
What is 'rhetorical decoupling' in this paper?
It refers to the finding that neutral agents' written justifications became semantically similar to the biased agents' text (measured with SBERT cosine similarity) even when their numeric opinions had shifted only moderately - in other words, the language converges faster than the actual stated positions, so text-only monitoring can overestimate consensus.
Is code or data available for this paper?
No. The paper describes the prompt template, model (Llama 3.2-8B), and simulation parameters in detail, but no code repository or dataset release is mentioned.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text contains garbled passages (e.g., 'numerical convergenumericalutral' in the abstract, 'simulating a discussiregulations' in Section 1), so some claim phrasing was reconstructed from context.
- Figures 3-5 (opinion trajectories and the alignment heatmap) cannot be analyzed from text; numeric judgments rely on Table 2 and in-text values.
- Figure 5's alignment values are never tabulated, so the magnitude of the claimed rhetorical/numerical decoupling could not be verified.
- Review is based on the arXiv v1 full-text extraction only; supplementary material, if any, was not seen.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.