9,455 LLM trajectories show neighbor data improves individual forecasts across 16 datasets while collective gains hinge on transfer conditions

The arXiv preprint by Igor Itkin tested whether compact surrogate models can reproduce collective behavior in LLM-agent societies. Researchers ran controlled experiments on opinion dynamics, explicitly limiting observation windows and comparing against simple baselines. Neighbor-augmented models consistently outperformed isolated predictors, yet collective fidelity only held under matched transfer conditions; mismatched history effects disappeared in 24 new statements. Qwen alone showed benefit from three-round history. These results underscore that individual-level gains do not automatically scale to group-level fidelity without direct collective validation. The work highlights a methodological gap: most prior LLM-society studies validated only at the agent level, leaving collective drift unmeasured. Direct collective testing plus explicit observation limits, as performed here, provide a clearer benchmark than agent-level metrics alone. Future studies should report both levels and include non-LLM baselines to isolate whether gains stem from architecture or from the added neighbor signal.