Large Language Models (LLMs) are constantly evolving, but how do we accurately measure their progress? Benchmarks like HumanEval are designed to test their coding abilities, but a lurking problem threatens to skew the results: data leakage. Imagine training for a race on the same track where the competition will be held—you'd have an unfair advantage. Similarly, if an LLM has already seen the benchmark problems during its training, its performance will appear inflated. Researchers are tackling this challenge head-on with a clever solution: combinatorial test design. Instead of using fixed problems, they create templates that can generate numerous variations. Think of it like a recipe with interchangeable ingredients. This approach allows for constant evolution of the benchmark, keeping it fresh and preventing LLMs from simply memorizing the answers. Initial experiments comparing HumanEval with a variant built using this method, HumanEval T, show promising results. Popular LLMs like GPT and Claude performed consistently lower on HumanEval T, suggesting data leakage might be a bigger problem than we thought. This research highlights the importance of robust evaluation methods in the rapidly advancing field of AI. As LLMs become more integrated into our lives, ensuring their true capabilities are accurately measured is crucial. The next step is expanding this research to other benchmarks and exploring new ways to measure the difficulty of these generated tasks. This is just the beginning of a critical conversation about how we evaluate—and truly understand—the progress of artificial intelligence.