Researchers have introduced FlavourBench, a benchmark for evaluating language models. Unlike traditional benchmarks, FlavourBench uses executable culinary ground truth provided by a versioned culinary system. The benchmark consists of tasks that ask models to suggest three-ingredient portfolios from eight ingredients, with all possible portfolios scored by Epicure before model execution. This approach eliminates differential missingness and allows for simultaneous evaluation of multiple models on an identical set of tasks. --- Why it matters: This matters because it provides a more robust way to evaluate language models, which can be prone to biases and inconsistencies. By using executable ground truth, FlavourBench enables researchers to compare models on the same tasks and metrics, leading to more reliable rankings and insights. Source: https://arxiv.org/abs/2608.20574