What is the Contagion Networks framework?

It is a mathematical framework that quantifies how systematic evaluation biases from large language models dynamically spread across multi-agent networks through inter-agent propagation matrices.

Why does evaluator bias propagation matter?

Unchecked bias can severely compromise system fairness and reliability, potentially causing widespread errors in automated decision-making pipelines and triggering a crisis of trust.

What practical strategy mitigates this bias propagation?

Expanding the evaluation committee from one to three agents reduces effective bias propagation by 72.4%, offering a robust and scalable architectural solution for deployment.

Contagion Networks：多智能體系統中評估者偏差的傳播機制研究

本文提出 Contagion Networks 框架，旨在量化大型語言模型作為評估者時，其系統性評估偏差在多智能體網路中的傳播規律。研究透過受控的三智能體實驗，利用 DeepSeek-chat 並設定結構化、平衡及基於證據三種偏差畫像，構建了跨智能體傳播矩陣 Gamma_3。結果顯示，即便在同一底層模型內部，評估偏差仍能持續傳播，傳播係數 gamma 介於 0.157 至 0.352 之間。研究識別出由譜半徑 rho(Gamma_N) 主導的三種傳播機制，並發現同質模型智能體的傳播係數僅為跨模型場景的三分之一至五分之一，處於抑制狀態。此外，將評估委員會規模從 k=1 擴大至 k=3 可有效降低 72.4% 的有效傳播，為緩解偏差提供了可操作的策略。作者已開源該實驗框架，推動多智能體系統評估的可靠性研究。

Sources

arXiv