How AI describes ZeroSumEval

ZeroSumEval is an extensible framework for evaluating large language models (LLMs) through competitive, multi-agent games that scale in difficulty as models improve. It pits models against each other in structured games with clear win conditions to assess knowledge, reasoning, planning, and other capabilities, using DSPy-based optimization to evaluate self-improvement while keeping competition fair. The project provides a growing evaluation suite and configuration for running these dynamic benchmarks rather than fixed, subjective assessments.