cs.MAApr 10, 2026
SaveAligned Agents, Biased Swarm: Measuring Bias Amplification in Multi-Agent Systems
Organizations: Shanghai Jiao Tong University · Shanghai Innovation Institute
Abstract
While Multi-Agent Systems (MAS) are increasingly deployed for complex workflows, their emergent properties-particularly the accumulation of bias-remain poorly understood. Because real-world MAS are too complex to analyze entirely, evaluating their ethical robustness requires first isolating their foundational mechanics. In this work, we conduct a baseline empirical study investigating how basic MAS topologies and feedback loops influence prejudice. Contrary to the assumption that multi-agent collaboration naturally dilutes bias, we hypothesize that structured workflows act as echo chambers, amplifying minor stochastic biases into systemic polarization. To evaluate this, we introduce Discrim-Eval-Open, an open-ended benchmark that bypasses individual model neutrality through forced comparative judgments across demographic groups. Analyzing bias cascades across various structures reveals that architectural sophistication frequently exacerbates bias rather than mitigating it. We observe systemic amplification even when isolated agents operate neutrally, and identify a 'Trigger Vulnerability' where injecting purely objective context drastically accelerates polarization. By stripping away advanced swarm complexity to study foundational dynamics, we establish a crucial baseline: structural complexity does not guarantee ethical robustness. Our code is available at https://github.com/weizhihao1/MAS-Bias.
Figures & tables
Figure 1: Two Parallel, Transformative Trends Shaping the Current AI Landscape. Left: The rapid advancement of powerful single-agent tooling, such as Codex ( OpenAI, 2026 ) and Claude Code ( Anthropic, 2026b ) , which excel in complex coding and generic problem-solving. Right: The paradigm shift towards complex Multi-Agent Systems (MAS), like Agent Teams ( Anthropic, 2026a ) and Agent Swarms ( Kimi, 2026 ) , designed for collaborative task execution. Together, these trends expose a critical core challenge: understanding how individual errors, uncertainties, and latent biases accumulate or mitigate when deployed in complex, real-world collaborative networks.
Figure 2: Overview of Our Framework for Investigating Iterative Bias Amplification in LLM-based MAS. Top: We propose Discrim-Eval-Open, an open-ended benchmark with multi-attribute options, to measure how an initial preference is progressively amplified as its reasoning is passed through a sequential chain of agents. Bottom: We then systematically evaluate whether common MAS architectures—employing diverse personas, specialized functions, complex topologies, and deeper iterations—can mitigate this fundamental amplification dynamic.
Figure 3: Demographic Distribution of Protagonist Profiles in Discrim-Eval-Open Dataset. The benchmark includes 210 unique profiles with a diverse spread of attributes. Left: Age distribution covers a wide spectrum from individuals in their 20s to over 100. Center: Gender distribution is perfectly balanced, with exactly 70 instances each for Male, Female, and Non-binary identities. Right: Race and ethnicity distribution is approximately balanced across five distinct groups. This balanced and diverse composition is designed to enable a robust and fair assessment of system-level bias across sensitive attributes.
Figure 4: Systematic Evaluation of MAS Architectures as Potential Mitigators of Iterative Bias Amplification. Left: We investigate agent specialization in linear chains, testing whether assigning diverse personas (e.g., Doctor, Lawyer) and functions (e.g., Analyst, Reflector) can introduce varied perspectives to curb the amplification effect. Right: We evaluate the role of communication structure by designing more complex topologies (Spindle, Parallel, Fully-Connected) and assess the impact of system depth by iterating the fully-connected unit. These configurations allow us to test if MAS architectural sophistication can overcome bias amplification.
Figure 5: Empirical Results Showing MAS Specialization Fails to Mitigate Iterative Bias Amplification. The plots show the relative Gini coefficient across four sequential agent layers (L1-L4) for eight different LLMs. (a) A baseline chain with identical roles confirms consistent amplification. Testing the mitigation hypothesis, we find that introducing (b) diverse personas, (c) specialized functions, or (d) a mix of both does not prevent the overall upward trend of bias amplification. Notably, while the ‘Reflector’ agent at L3 in panel (c) provides a partial and temporary reduction in bias for some models, the amplification trend consistently resumes by the final layer.
Figure 6: MAS Architectural Complexity Fails to Mitigate but Exacerbate Bias Amplification. These plots show that complex communication structures and increased system depth do not solve the core issue of iterative amplification. (a-c) Bias progressively amplifies across all tested four-layer topologies (Spindle, Parallel, and Fully-Connected). (d) Furthermore, increasing system depth by iterating a fully-connected unit end-to-end (from I0 to I4) results in a particularly steep and sustained amplification of bias. These findings demonstrate that neither sophisticated information flow nor deeper systems in MAS are effective mitigators.
Figure 7: A Neutral Trigger Can Initiate a Cascade of Bias Amplification, Revealing System Fragility. This qualitative example compares two scenarios using a sequential MAS of Gemini 3.1 Pro ( DeepMind, 2026 ) agents. Top Path: Without external input, the well-aligned system maintains a balanced and fair probability distribution. Bottom Path: However, introducing a single, seemingly objective sentence acts as a trigger, creating an initial bias that is then rapidly and progressively amplified by subsequent agents. This highlights a critical vulnerability: MAS are susceptible to having latent biases triggered and amplified by external context.
Figure 8: MAS Tendency toward Favoring Younger Individuals, Women, and Black Communities. Results are derived from the whole benchmark across 70 scenarios, in a four-layer sequential MAS composed of DeepSeek-V3.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Impact of Historical Information on Bias Amplification. MAS is constructed by sequentially connecting four agents using either DeepSeek-V3 or DeepSeek-R1. In the left subfigure, each agent receives the accumulated viewpoints from all preceding agents, whereas in the right subfigure, each agent only receives the opinion of its immediate predecessor. Results show that bias is progressively amplified in both settings, with more pronounced amplification observed when agents are exposed to a greater amount of historical context.
Figure 10: Impact of Mixed Personas and Functions on Bias Amplification in MAS Construction. A four-agent MAS is constructed with a hybrid configuration: Agent 1 (left) serves as a judger, Agent 2 (top) as a doctor, Agent 3 (bottom) as an engineer, and Agent 4 (right) as a summarizer. Different LLMs are used to instantiate the agents, and variance is employed as the metric to quantify bias. Results show a clear trend of progressive bias amplification across the agent chain.
Figure 11: Effect of Spindle MAS Topology on Bias Amplification, Measured by Variance. Agents 1–7 represent Judger, Doctor, Engineer, Summarizer, Lawyer, Merchant, and Summarizer, respectively. Lighter colors indicate higher variance, corresponding to more extreme bias. Results across multiple MAS configurations using different LLMs show that bias is progressively amplified, particularly between key functional nodes: Agent 1, Agent 4, and Agent 7.
Figure 12: Impact of Iteration Rounds on Bias Amplification in MAS. The MAS is constructed using the same LLM across all nodes, with a topology consisting of four sequentially connected fully-connected sub-units. Higher variance indicates more extreme bias. The dashed baseline denotes the output of the first node (Judger) in the first sub-unit, while the solid lines represent the outputs of the final Summarizer node in each sub-unit. Results demonstrate that bias is progressively amplified over successive iteration rounds.
| Model | Dataset | Biased | All Cases | Model | Dataset | Biased | All Cases |
| DeepSeek-V3 | Explicit | 820 | 9450 | GPT-4o | Explicit | 981 | 9450 |
| DeepSeek-V3 | Implicit | 942 | 9450 | GPT-4o | Implicit | 1072 | 9450 |
Table 4: Current LLMs Exhibit Limited Detectable Bias on Tamkin et al. (2023) .
More in this topic
Algorithmic Bias
Algorithmic bias studies how automated systems produce or amplify unfair outcomes across demographic groups, and how to measure and reduce them.
See all 100 papers