CDBench: Benchmarking the mutation testing capabilities of LLMs with code defenders

Abstract Most traditional benchmarks for evaluating Large Language Models (LLMs) in software development suffer from a narrow focus, high risks of data contamination, and static difficulty levels that fail to keep pace with rapid model evolution. To address these limitations, we introduce CDBench, a novel zero-sum benchmark based on the Code Defenders mutation testing game. By pitting models against each other in a competitive environment—where “attackers” introduce code mutations and “defenders” create tests to detect them—CDBench establishes a dynamic difficulty curve that scales naturally without human intervention. Our experiments reveal that while LLMs can generate diverse mutations, they often struggle with code validity; nevertheless, the framework effectively distinguishes model capabilities, highlighting the superior test generation of models like Gemini 2.5 Pro while exposing the instruction-following limitations of reasoning models. These findings demonstrate that zero-sum games offer a viable, contamination-resistant solution to the stagnation of current evaluation methodologies.