OpenIIR lets researchers run hundreds of LLM personas in IR simulations

Saber Zerhoudi introduces OpenIIR, an open simulation platform designed specifically for information retrieval research. It enables researchers to run experiments with hundreds of LLM-driven personas in a highly parameterized, reproducible manner. The platform supports four distinct types of multi-agent studies: deliberative panels, social media platforms, curated recommender feeds, and an evolutionary co-evolution setup between content producers and credibility detectors. Each study can be configured with tunable priors, rounds, constraints, persona budgets, retrieval policies, ranker choices, intervention timings, and mutation rates. This allows researchers to compare outcomes side by side under different settings.

Every simulation run produces structured outputs such as argument graphs, exposure logs, fitness traces, and transcripts, which downstream evaluators can consume directly. A new study is implemented as a 200–400 line plug-in that builds on a shared core consisting of an agent runtime, world-model store, retrieval primitives, claim extractor, and persona ontology. The paper’s contributions include the shared core, a type interface for pluggable scenarios, four released types with reference runs, and six modular extensions sketched against open IR research questions. OpenIIR is designed to advance reproducible and scalable IR experimentation, particularly for studying algorithm-mediated information ecosystems.