← News

Agents & MCPSources: arXiv · 18 Nov 2026

arXiv study links multi-agent network shape to how bias spreads through LLM systems

A controlled bias injection study finds sequential agent networks diverge most when fed biased responses, while fully connected and mesh structures resist more.

In plain words

Researchers tested what happens when one AI agent in a multi-agent system is fed a biased answer. They found the effect depends heavily on how the agents are connected: agents arranged in a chain passed the bias along more than densely connected groups. This suggests auditing such systems should account for their network structure.

What was announced

The arXiv paper introduces a controlled bias injection framework that replaces selected agent responses with predefined biased ones, allowing intervention-based analysis of how network topology and injection configuration shape system behavior. The authors argue that fairness risks arising from interactions among components can go undetected when auditing components or collective outcomes in isolation.

Using decision divergence metrics at the agent, network, and system levels, the study evaluated five topologies on three binary classification tasks with Qwen and Gemma family models. Sequential networks showed the greatest divergence, while fully connected and mesh networks showed less. In star networks, injecting the central agent had larger effects than injecting peripheral agents. The models also differed in whether divergence followed the injected bias. The authors present this as motivation for topology-aware fairness auditing.

What it does not show

  • The evidence is limited to the paper's abstract; full methodology, task details, and quantitative results are not available in the supplied passages.
  • Findings cover five topologies, three binary classification tasks, and Qwen and Gemma models only; generalization to other structures, tasks, or model families is not established by the abstract.
  • No independent replication of the results is indicated in the supplied evidence.

For security and IT teams

Scope
The study concerns LLM-based multi-agent systems in classification settings and does not, per the abstract, address production deployment guidance.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.