Local Predictability and Collective Fidelity in LLM-Agent Societies
Abstract
Compact surrogates could reduce the cost of simulating large language model societies, but must reproduce collective behavior. We compare individual predictions and collective forecasts using 9,455 published trajectories and new experiments on opinion dynamics. Neighbor information improves individual prediction in all 16 public-data settings and pooled collective forecasts on held-out questions, although collective gains depend on transfer conditions. Tests on 24 new statements do not confirm earlier contrasting history effects in forecasts from the initial state. Qwen benefits from history after three observed rounds. These findings motivate direct collective validation, explicit limits on available observations, and comparisons with simple baselines.
Related
- WolfSociety: Understanding Collective Risk from Harmful-Agent Scaling in Financial Agent Societies
- Benchmarking large language model agent societies against human behavioural distributions
- Atom as an Agent: Emergent Order by Collective Predictive Coding
- Collective impact of opinion formation on agent-based epidemic dynamics
- Role differentiation as ignition of a collective information engine: Structuration in Agent Populations
BibTeX
@misc{itkin2026local,
title = {{Local Predictability and Collective Fidelity in LLM-Agent Societies}},
author = {Igor Itkin},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.35813},
eprint = {2609.35813},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.35813},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Local Predictability and Collective Fidelity in LLM-Agent Societies
AU - Igor Itkin
PY - 2026
DA - 2026/09/20
PB - arXiv
DO - 10.48550/arxiv.2609.35813
UR - https://arxiv.org/abs/2609.35813
AB - Compact surrogates could reduce the cost of simulating large language model societies, but must reproduce collective behavior. We compare individual predictions and collective forecasts using 9,455 published trajectories and new experiments on opinion dynamics. Neighbor information improves individual prediction in all 16 public-data settings and pooled collective forecasts on held-out questions, although collective gains depend on transfer conditions. Tests on 24 new statements do not confirm earlier contrasting history effects in forecasts from the initial state. Qwen benefits from history after three observed rounds. These findings motivate direct collective validation, explicit limits on available observations, and comparisons with simple baselines.
ER -