MiroBench: Benchmarking Realism in Agentic Simulation of Real-world Discussions

Researchers have introduced MiroBench, a new benchmark built from 4,292 real Reddit threads to measure how faithfully large language model agents simulate the distributional patterns and interaction dynamics of real online discussions [1][2]. The benchmark, described in a paper submitted to arXiv on 10 May 2026, addresses a gap in evaluating LLM-based social simulation. While LLM agents are increasingly deployed to mimic real-world interactions, existing evaluations remain fragmented, making it difficult to compare systems or track progress [1][2]. The authors selected Reddit threads as a testbed because they provide public, topic-grounded, multi-party exchanges where users share experiences, debate, seek advice, and express emotion [1][2]. MiroBench applies statistical tests to compare generated and real discussions across four dimensions: repetition and semantic uniformity, narrative content, toxicity and aggression, and structural complexity [1][2]. Experiments spanning five domains and five models found that current simulators remain distributionally mismatched with authentic Reddit threads [1][2]. A lightweight prompt-based improvement procedure yielded only limited gains, underscoring the difficulty of replicating genuine community dynamics [1][2]. Large language models are machine learning systems trained on vast text corpora for natural language generation [8]. arXiv, where the MiroBench paper appears, is an open-access repository of scientific preprints that has hosted over two million articles since its launch in 1991 and now receives roughly 24,000 submissions per month [6]. The platform also supports community-developed tools through arXivLabs, a framework that allows collaborators to build features such as citation explorers and code finders directly on article pages [4][5]. The MiroBench authors frame their work as a concrete step toward measuring, diagnosing, and improving realism in LLM-based social simulation [1][2]. By providing a standardized set of statistical comparisons, the benchmark aims to help researchers identify where synthetic discussions diverge from human behavior and to track incremental progress over time [1][2].