Model Cross-Contamination
Unintended behavioral influence between agents powered by different foundation model vendors.
Model cross-contamination is the phenomenon where agents powered by different foundation model vendors influence each other's behavior through shared interaction in a multi-agent environment, even when each agent's model is fixed and isolated. The contamination occurs through the exchange of information — conversation content, observed behaviors, and environmental outcomes — that causes agents to adopt strategies, reasoning patterns, or preferences from other models without being explicitly told to do so.
The mechanism is distinct from simple imitation: agents do not merely copy outputs, they update their internal reasoning based on what they observe other agents saying and doing. In heterogeneous populations where agents from different vendors share a world, this creates a form of behavioral mixing that can erode the intentional differences between model families. A Claude-powered agent, for instance, might begin exhibiting planning heuristics it observed from a GPT-powered agent, even though both were given identical initial instructions. The effect compounds over long time horizons as agents adapt to each other's presence.
The primary risk is loss of vendor independence in multi-agent evaluations. If agents using different models converge behaviorally, cross-vendor studies become less meaningful — observed differences diminish not because of inherent model capabilities but because of mutual contamination. This also complicates safety evaluation: a model that appears safe in isolation may behave differently when exposed to the reasoning patterns of a less cautious model in a shared environment. Mitigations include information barriers, read-only observation modes, and controlled interaction protocols, but these constrain the realism that makes multi-agent simulations valuable.
It remains an open question whether cross-contamination is a property of the interaction protocol, the models themselves, or the specific tasks being performed. Whether contamination effects are symmetric — whether a more capable model influences a less capable one more than the reverse — is also unknown. The phenomenon raises fundamental questions about what it means for a model to have a consistent identity when its behavior can be shaped by its social environment.