Source-linked AI summary
Bias Amplification in Multi-Agent Network: How Biased Agents Shape Opinions and Rhetoric
Omran Berjawi, Giuseppe Fenza, Rida Khatoun
TL;DR
The paper investigates how bias propagates between LLM agents when a persistent minority interacts with otherwise neutral agents. Through controlled multi-topic simulations and comparison with the FJ model, it finds stronger and faster opinion shifts in Llama 3.2, alongside rhetorical alignment that can outpace numerical convergence.
Problem
Evidence remains limited on how a persistently biased minority affects neutral LLM populations and how bias propagates through language as well as numerical opinions.
Method
The study varies the size of a fixed-stance biased minority in controlled multi-topic Llama 3.2 agent simulations while tracking opinions and textual justifications against the FJ model.
Results
Llama 3.2 interactions produce stronger and faster convergence toward biased positions than the FJ baseline, while rhetorical alignment increases with exposure even when opinion shifts remain moderate.
Takeaways & Limitations
Bias in LLM-agent networks develops through both numerical opinion change and increasingly homogeneous language, making apparent linguistic consensus potentially coexist with meaningful disagreement.
Takeaways & Limitations
The findings come from Llama 3.2-8B with one agentic framework, so generalization across models and frameworks remains unknown.
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly deployed in applications involving interaction between agents, where their output plays a role in collective reasoning and decision-making processes. Despite significant research into the functioning of LLMs in such multi-agent systems, the processes of bias propagation in such systems are still a challenge. This work studies how biased opinions are propagated in the form of textual interaction in an environment of LLMs, in which a minority of agents maintain persistent extreme opinions, while the remaining agents iteratively update their beliefs through structured textual interactions. The findings show that even the presence of a small percentage of biased agents in such a system leads to significant shifts in the opinions of non-biased agents. It suggests that for the same percentage of biased agents, the shifts occur more quickly for the Llama~3.2 model when compared to a classical Friedkin-Johnsen (FJ) model. Further semantic analysis demonstrates that rhetorical consistency in textual explanations increases systematically with biased exposure and, importantly, is partially decoupled from numerical convergenumericalutral agents adopt the vocabulary employed by the biased agents even in configurations where their numerical opinion shifts remain moderate. The research helps explain how bias and language develop together in multi-agent language model ecosystems.
1 Introduction
This paper examines how persistent biased minorities influence neutral LLM agents through repeated textual interaction, affecting both opinions and rhetoric. It proposes controlled simulations and comparison with the classical model to study these effects.
- 1 Introduction: The motivation is that LLM agents increasingly participate in decision-making and public discourse while inheriting biases from training data and societal assumptions.Understanding propagation during interaction is presented as important for responsible development and AI-mediated communication.
- 1 Introduction: Persistent biased minorities can shift neutral agents’ opinions and rhetorical patterns through repeated LLM interactions.The study tests whether exposure produces directional opinion change and linguistic alignment.
- 1 Introduction: The experiments simulate fully connected LLM-agent discussions across four controversial topics, with biased agents holding fixed extreme views and neutral agents updating gradually.Analysis tracks both opinion spreading and the evolution of textual justifications.
- 1 Introduction: The paper contributes a controlled framework, a comparison between LLM interactions and a classical model, and joint analysis of numerical opinions and language.These contributions target how quickly opinions settle and how language aligns with bias over time.
2 Related Work
Prior work models polarization and opinion formation with classical and LLM-based systems, but leaves controlled minority-bias propagation and linguistic influence insufficiently examined. This paper addresses both gaps with matched benchmarking and dual-layer analysis.
- 2 Related Work: Earlier studies examine polarization, echo chambers, homophily, social influence, and opinion development in classical and LLM-based simulations.LLM frameworks cover multi-topic populations, social-media events, and heterogeneous agent or community structures.
- 2 Related Work: Prior evidence reports that repeated LLM interaction can increase bias, while prompt design, network homophily, and population composition affect collective outcomes.Other studies also compare LLM simulations with classical models and evaluate their agreement with survey-based opinion data.
- 2 Related Work: Existing work does not isolate the causal effect of a persistently biased minority on an otherwise neutral population across controlled exposure levels.Earlier studies focus on heterogeneous populations or endogenous model biases rather than graded minority exposure.
- 2 Related Work: Prior studies track bias propagation numerically, whereas this framework jointly measures opinion values and semantic alignment in textual justifications.It varies fixed-stance minority size, benchmarks against the FJ model, and examines rhetorical-opinion decoupling.
3 Methodology
The methodology simulates repeated opinion and justification updates in a fully connected LLM-agent network containing neutral and persistent biased agents. Structured prompts, categorical opinions, parsing, and bounded updates support comparison of numerical and textual dynamics.
- 3.1 Agent Network and Interaction Structure: The framework uses repeated interactions on a fixed network, partitioning agents into updating neutral agents and persistent biased agents with fixed extreme stances.This design exposes neutral agents to the extreme biased opinion at each round.
- 3.2 Agent Initialization: Each agent holds a categorical opinion in {−2, −1, 0, 1, 2} and generates a textual justification for its stance.The categories range from Strongly Disagree to Strongly Agree, with extreme values representing maximal disagreement or agreement.
- 3.2 Agent Initialization: Neutral agents receive moderate initial opinions from {−1, 0, 1}, while biased agents are initialized at +2 to isolate sustained one-sided minority influence.The Ab = 0 configuration captures drift attributable to initialization alone.
- 3.2 Agent Initialization: At each round, structured prompts provide prior beliefs, social opinions, and justifications, requiring an updated categorical stance and short explanation.The prompt also limits opinion movement to at most one categorical step per round.
- Opinion Update Mechanism: The pipeline maps LLM-generated stances to numeric values, applies bounded-step updates, and records opinions and justifications in JSON.Neutral agents update, whereas biased agents keep their numerical opinion fixed but regenerate justifications each round.
4 Experimental Setup
The experiments vary biased-minority exposure across four policy topics and compare language-mediated Llama 3.2 dynamics with a matched classical FJ baseline. They quantify numerical opinion, convergence, and textual semantic alignment over repeated simulations.
- Discussion Topics: The study covers AI safety regulation, vaccine mandates, immigration policy, and climate change to test whether amplification extends across policy domains.The biased stances favor stronger regulation, vaccine mandates, stricter immigration policies, and aggressive climate intervention, respectively.
- Network and Simulation Parameters: 50 agents interact on a fully connected network, with biased-agent counts Ab ∈ {0, 2, 4, 6, 8, 10}, spanning 0% to 20% of the population.Ab = 0 is the fully neutral baseline; higher values introduce progressively stronger minority influence.
- LLM Simulation: Llama 3.2 agents use structured textual interaction, generating numerical opinions and textual justifications with identical temperature-0.7 settings across rounds.Each agent is implemented independently with LangChain, and semantic alignment compares neutral-agent justifications with biased-agent justifications using SBERT embeddings and cosine similarity.
- Baseline Comparison: The FJ baseline combines neighbor influence with intrinsic initial beliefs while matching population, network structure, bias levels, and stubbornness assumptions to the Llama 3.2 condition.Biased agents are fully stubborn at +2, while neutral agents use λn = 0.5 to balance individual conviction and social influence.
- Evaluation Procedure: Each simulation runs for K = 10 rounds and is repeated 15 times, with findings reported as means and 95% confidence intervals from standard errors.The only experimental difference between conditions is neutral-agent opinion updating: numerical averaging versus language-mediated reasoning.
- Evaluation Metrics: Bias amplification is evaluated over neutral agents using mean opinion, mean opinion shift, distance to the biased anchor, and semantic alignment of textual justifications.These measures capture directional stance, movement from initial opinions, convergence toward the imposed extreme, and rhetorical similarity, respectively.
5 Experimental Results
Across topics, increasing exposure to biased agents shifts Llama 3.2 opinions more strongly and quickly than in the FJ model, while rhetorical alignment rises even when numerical shifts remain moderate.
- Temporal Evolution of Opinion Trajectories: Llama 3.2 opinion shifts increase with biased-agent count across all topics, with most change occurring within the first three rounds before stabilizing.At Ab = 8 and 10, mean opinions are highest across TReg, TV ac, TImm, and TCLIM.
- Aggregate Bias Metrics Across Topics: The FJ model changes gradually and nearly linearly, with minimal differences between bias levels and a shift approaching 0.5 at Ab = 10.Its distance to the skewed reference decreases from 2.05 to 1.53 as bias increases.
- Aggregate Bias Metrics Across Topics: Llama 3.2 produces larger shifts than FJ across topics, reaching a mean shift of 1.49 for AI Safety at Ab = 10 versus 0.02 at Ab = 0.For AI Safety, distance to the anchor falls from 1.97 to 0.50; analogous patterns occur for the other topics.
- Aggregate Bias Metrics Across Topics: Vaccine mandates show a mean shift of 0.66 even at Ab = 0, indicating model-level directional bias that becomes stronger when biased agents are present.The passage characterizes language-based interactions as amplifying minority effects beyond classical-model behavior.
- 5.2 Affective Alignment Analysis: Semantic alignment rises monotonically with biased-agent count across all topics, reaching its highest levels at Ab = 10.TReg has the highest alignment at larger bias levels, while TCLIM remains comparatively lower.
- 5.2 Affective Alignment Analysis: Alignment begins increasing substantially at Ab = 4 and 6 even where numerical opinion shifts remain moderate, showing that language homogenizes faster than opinions converge.Figure 5 reports one final-round alignment value for each of four topics and six bias levels, with interpolated visualization between measured cells.
6 Discussion
The discussion concludes that persistent biased minorities systematically shift neutral agents’ opinions and increasingly shape their rhetoric. These effects are stronger in the Llama 3.2 setting, where linguistic convergence can outpace numerical convergence.
- H1: Opinion Shifts: Exposure to a persistently biased minority produces measurable directional drift in neutral agents even at Ab = 2, increasing as more biased agents are added.Both rising mean opinion and shrinking distance to the biased anchor support this pattern.
- H1: Opinion Shifts: Llama 3.2 shows faster and larger shifts than the FJ baseline, consistent with stronger minority influence under language-mediated interaction.The discussion associates this difference with the richer rhetorical context exchanged by agents.
- H2: Rhetorical Alignment: Neutral agents’ justifications increasingly resemble the biased minority’s rhetoric across topics, even when their numerical opinions remain only moderately shifted.The result suggests that agents adopt biased rhetoric more extensively than the biased position itself.
- H2: Rhetorical Alignment: Because language homogenizes while meaningful opinion disagreement persists, text-only monitoring can make the system appear more consensual than it is.This asymmetry partially masks cognitive diversity.
7 Limitations
The study is a proof-of-concept with important scope and measurement limitations that leave replication, robustness, temporal precedence, and directional symmetry unresolved.
- The experiments use only Llama 3.2-8B with LangChain, so whether the rhetorical–numerical convergence decoupling generalizes across models or agentic frameworks remains unknown.The authors identify replication across model families and alternative frameworks as a priority for future work.
- The bounded-step instruction and agent-role framing may shape responses, but systematic prompt ablations were not conducted.Future ablations are needed to establish robustness to instruction variations.
- Final-round semantic alignment measurements show differing convergence magnitudes but cannot establish whether rhetorical alignment precedes individual opinion change.Per-round alignment tracking and lagged analyses are required to test temporal precedence.
- Because biased agents were anchored only at +2, the measured amplification may interact with an intrinsic model prior toward positive positions.Testing −2 anchors would separate promoted-stance effects from underlying model bias and assess symmetry.
- Overall, the paper establishes measurable, directional minority bias amplification but leaves important questions open as a proof-of-concept.
8 Conclusion
The paper studies how persistent biased agents shape both opinions and textual justifications in multi-agent LLM networks. Controlled multi-topic simulations show stronger and faster convergence toward biased positions than the FJ baseline, while semantic alignment rises even when discrete opinions retain diversity.
- Llama 3.2 interactions produce stronger and faster convergence toward biased positions than the Friedkin–Johnsen baseline.
- Increasing exposure to biased agents causes systematic and nonlinear opinion shifts among neutral agents across multiple topics.
- Semantic alignment in textual justifications increases consistently with biased exposure, even when some diversity in discrete opinions remains.
- The findings highlight that LLM interactions shape both agents’ beliefs and their expression, raising considerations for socially sensitive AI deployments.