Contagion Networks: How Evaluator Bias Spreads in Multi-Agent Systems

Background and Context

In the rapidly expanding landscape of multi-agent systems, the integration of large language models as evaluative agents introduces a critical vulnerability: the propagation of systematic evaluation bias. When large language models are assigned the role of judges or reviewers within a network, their inherent biases do not remain isolated; instead, they interact and spread across agent boundaries, potentially compromising the fairness and reliability of the entire system. This phenomenon poses a significant challenge to the robustness of automated decision-making pipelines, where the integrity of the output is contingent upon the accuracy and impartiality of the evaluation process. The Contagion Networks framework has been introduced to address this gap by providing a formalized mathematical model to quantify and analyze how these biases diffuse through interacting networks of language models.

The core objective of this research is to establish a rigorous methodology for measuring the dynamics of bias transmission. By treating the multi-agent system as a network of nodes, the study aims to map the flow of evaluative distortions from one agent to another. This approach moves beyond static assessments of individual model performance to a dynamic analysis of systemic interactions. The framework is designed to identify the pathways and intensities through which bias travels, offering developers a visual and analytical tool to pinpoint vulnerabilities within their architectures. Understanding these mechanisms is essential for ensuring that multi-agent collaborations remain robust against information pollution and systematic errors that can arise from biased evaluations.

Deep Analysis

To empirically validate the Contagion Networks framework, the researchers designed a controlled experimental environment featuring three intelligent agents, all powered by the DeepSeek-chat model. This homogenous setup allowed the team to isolate the effects of bias profiles from the confounding variables of different model architectures. The agents were configured with three distinct bias profiles: structured bias, balanced bias, and evidence-based bias. By manipulating these profiles while keeping the underlying model constant, the study could precisely observe how specific types of bias influence propagation behavior. The central analytical tool in this experiment was the inter-agent propagation matrix, denoted as Gamma_3, which captures the probabilistic transfer of bias between any two agents in the network.

The analysis of the Gamma_3 matrix revealed that evaluation biases persist and spread even among agents sharing the exact same underlying model. The propagation coefficients, represented by gamma, were found to range between 0.157 and 0.352. This quantitative evidence confirms that bias transmission is an intrinsic property of the interaction dynamics, not merely a byproduct of architectural differences. The study identified three distinct propagation mechanisms governed by the spectral radius, rho(Gamma_N): suppression, criticality, and explosion. These mechanisms describe the long-term behavior of bias in the network, determining whether the distortion will die out, stabilize, or amplify uncontrollably over successive interactions.

A key finding of the deep analysis is the significant difference in propagation strength between homogenous and heterogeneous agent configurations. The experiments demonstrated that homogenous model agents exhibit propagation coefficients that are only one-third to one-fifth of those observed in cross-model scenarios. In previous work, such as the MM-EPC framework, cross-model propagation coefficients ranged from 0.85 to 1.3, indicating a much higher risk of bias amplification. The lower coefficients in the homogenous setting suggest that model similarity acts as a buffer, keeping the system in a suppressed state where bias does not easily escalate. This insight highlights the importance of model diversity in managing bias propagation risks, as sharing the same underlying architecture may inadvertently create predictable patterns of bias transfer that are easier to mitigate than the chaotic spread seen in heterogeneous networks.

Industry Impact

The implications of the Contagion Networks framework extend significantly to the industrial deployment of multi-agent systems, particularly in high-stakes applications such as automated code review, content moderation, and complex decision support systems. In these domains, the unchecked spread of evaluation bias can lead to systematic errors that scale rapidly, eroding trust in AI-driven processes. For instance, if a biased evaluator agent consistently downgrades the quality of code generated by another agent, this bias can cascade through the network, leading to the rejection of valid solutions or the acceptance of flawed ones. The Contagion Networks framework provides a standardized tool for developers to detect and mitigate these risks before deployment, ensuring that their systems are resilient to such cascading failures.

Furthermore, the research offers actionable strategies for designing more reliable multi-agent architectures. The finding that expanding the evaluation committee from a single agent (k=1) to a group of three agents (k=3) reduces effective propagation by 72.4% is particularly impactful. This substantial reduction demonstrates the power of collective decision-making in filtering out individual biases. For industry practitioners, this suggests that increasing the diversity and number of evaluators is a highly effective mitigation strategy. By implementing a committee-based evaluation mechanism, organizations can significantly lower the risk of bias propagation, thereby enhancing the overall reliability and fairness of their automated systems.

The open-sourcing of the experimental framework by the authors further amplifies the industry impact. It provides the broader developer community with a replicable and transparent tool for auditing their own multi-agent systems. This accessibility encourages the adoption of best practices in bias mitigation and fosters a culture of accountability in AI development. As multi-agent systems become more prevalent in critical infrastructure and business operations, the ability to quantify and control bias propagation will be a key differentiator between robust, trustworthy systems and those prone to systemic failure. The Contagion Networks framework thus serves as a foundational resource for building the next generation of reliable and fair AI systems.

Outlook

Looking ahead, the Contagion Networks framework opens new avenues for research into the alignment, fairness, and safety of multi-agent systems. The ability to quantify bias propagation provides a concrete metric for evaluating the social impact of AI agents, moving the discourse beyond abstract ethical principles to measurable technical parameters. Future research can build upon this foundation to explore more complex network topologies, larger agent populations, and dynamic bias profiles that evolve over time. Additionally, the insights gained from this study can inform the development of new training strategies that explicitly penalize bias propagation, encouraging models to act as neutral evaluators rather than carriers of distortion.

The discovery that homogenous models exhibit suppressed bias propagation suggests that architectural design plays a crucial role in bias management. Future work could investigate whether introducing controlled heterogeneity, such as combining models with different training data or objectives, can further reduce bias spread while maintaining performance. Moreover, the significant reduction in propagation achieved by expanding the evaluation committee to k=3 invites exploration of optimal committee sizes and compositions. Determining the point of diminishing returns for committee size could lead to more efficient and cost-effective evaluation protocols.

Ultimately, the Contagion Networks framework represents a significant step toward ensuring the reliability of multi-agent systems. By providing a rigorous mathematical and empirical basis for understanding bias propagation, it empowers developers to build systems that are not only intelligent but also fair and robust. As AI systems continue to integrate into critical decision-making processes, the principles outlined in this research will be essential for maintaining public trust and ensuring that the benefits of AI are distributed equitably. The open-source nature of the framework further ensures that these advancements can be rapidly adopted and improved upon by the global community, accelerating the development of safer and more reliable AI technologies.