The Metanym Game: A Self-Contained, Self-Consistent LLM Peer-Community Benchmark for Structural Intelligence

The article introduces the Metanym Game, a novel benchmarking framework for large language models (LLMs) that evaluates structural intelligence through a competitive word game format. This benchmark is designed to be contamination-resistant, allowing LLMs to generate and assess content without pre-existing test sets, and utilizes a unique spectral solution for measuring factual accuracy based on the consistency of peer ratings. The findings indicate a high correlation (Pearson r = 0.92) with established benchmarks like GPQA Diamond, highlighting the potential for this framework to provide a self-sustaining and evolving measure of LLM performance in generating and judging factual content.