Aumann-Shapley Attribution Scales to Million-Agent LLM Systems

A new technique enhances Aumann-Shapley path-integral attribution specifically for multi-agent systems (MAS) that utilize large language models (LLMs), scaling up to one million agents. Traditional attribution methods struggle with scalability, as they can only handle up to 1,000 agents, while important social dynamics like polarization and market panics occur with millions of participants. This innovative method complies with all four attribution axioms and is four to five times faster than the sampled Shapley method on similar hardware. It was tested using 14 days of public Bluesky data, involving 1,671,587 active users, allowing the researchers to conduct full-scale attribution and compare it with visibility-biased samples. This study connects theoretical attribution concepts with real-world applications in LLM-driven social simulations.