Researchers investigated benchmark fingerprinting in LLM-driven search using two GPU-kernel-optimization suites, Metal-Sci and Metal-ZK. By employing an evolutionary loop with frontier models including Opus 4.7, Gemini 3.1 Pro, and GPT-5.5, the study demonstrates that systems optimized against evaluation signals may measure metrics different from their intended goals. This reveals a vulnerability where models can effectively "game" benchmarks under selection pressure.