When Does Group Structure Emerge? Validating History-Dependent Group Behavior in LLM Agent Populations

Many important properties of multi-agent systems arise from interaction rather than from any agent in isolation. Group structure is one example: repeated interactions may cause agents to treat members of the same category differently, generalize past experience to unfamiliar group members, or carry group-level expectations across contexts. Yet apparent group bias can also be produced directly by a visible label or by individual reputation, making it difficult to determine when a genuinely history-dependent population-level pattern has emerged. We propose a construct-validity framework for measuring emergent group-conditioned behavior in LLM agent populations. The key comparison pairs a history-carrying agent with a matched copy that faces the same current decision and group cue without the relevant interaction history. Hidden-label, memory-scrubbing, novel-member, and reassignment probes further separate group-history effects from cue following, reciprocity, and re-identification. In a pilot with GPT-4.1-mini, immediate label effects are large and most apparent persistence among familiar partners can be explained by individual reciprocity. However, group-structured conflict changes how agents treat previously unseen members of a category, suggesting that population history can generalize beyond individual relationships. We use this framework to motivate broader questions about how group structure forms, spreads, and can be controlled in persistent LLM-agent populations.