Researchers have developed GRALIS, a new method for explaining how deep neural networks make decisions. GRALIS combines two existing methods, coalition-based and gradient-based attribution, into one estimator. This fusion provides guarantees that neither method can offer alone: it can accurately calculate the contribution of individual features to the model's output, and it can do so with a finite amount of data. The new method is based on a mathematical representation theorem that shows how any additive attribution functional can be represented in a unique way. Preliminary experiments show promising results, but more extensive validation is needed. --- Why it matters: This matters because current methods for explaining neural network decisions have limitations and inconsistencies, making it hard to compare their performance. GRALIS addresses these issues by providing a unified framework that combines the strengths of different attribution methods. Source: https://arxiv.org/abs/2605.05480