The Contagion Tensor: A Framework for Measuring Output-Distribution Coupling in Multi-Agent LLM Systems -- and Auditing the Claims It Enables

Researchers have proposed the Contagion Tensor, a framework designed to quantify how output distributions of large language models couple across agents, modalities, and time steps in multi-agent systems, according to a preprint published on arXiv [1]. The framework derives the Coupling Amplification Factor (CAF), a family of ratio-based metrics expressed as CAF = E[T_condition] / E[T_baseline], which provides unitless, baseline-referenced measurements with bootstrap confidence intervals [1]. The authors instantiated CAF in four variants and evaluated the strongest using a complete 2x2x2 block-orthogonal simulation design with modality-specific ablation [1]. The simulation initially showed an apparent image-condition super-linear effect with a CAF of 1.40 [1]. When the image perturbation module was disabled, however, the effect collapsed to a sub-linear CAF of 0.87, a shift of -0.53, while text conditions remained unaffected [1]. The study was supplemented with real-API experiments across two model families: DeepSeek-Chat, using 30 models, and GPT-4o-mini, using 15 models with real vision [1]. Under uniform personas, text-only communication produced a CAF of approximately 1.0 in both model families [1]. When diverse personas were introduced, the CAF dropped to 0.88, indicating convergence [1]. A within-model comparison on GPT-4o-mini showed a text-condition CAF of 1.02, while the real-vision condition yielded a CAF of 1.72, with a bootstrap confidence interval of [1.700, 1.733] and a delta of +0.70, which the authors said validates the simulation's super-linear image-condition prediction [1]. The preprint states that of 11 experimental conditions, 5 have been tested on real APIs and 6 remain unverified [1]. The authors describe their contribution as two-layered: a measurement instrument that makes output-distribution coupling quantitatively falsifiable, and a transferable ablation protocol that any modular multi-agent simulator can adopt to distinguish genuine coupling from design artifacts [1]. The work was posted on arXiv, a preprint server operated by Cornell University that hosts papers across disciplines including computer science and machine learning [2].