Apple AI Updates: July 28, 2026

1. Apple’s GH-ESD Surfaces Systematic Failures in Instance-Level Vision Models

Apple researchers introduced GH-ESD (Grounded Hypothesis-Driven Error Slice Discovery), a framework for finding semantically coherent subsets where vision models consistently fail on instance-level tasks such as object detection and segmentation. The authors note that prior error slice methods target image-level classification through representation clustering or predefined attributes, which does not capture the contextual, relational, and spatially grounded patterns behind instance-level failures. GH-ESD uses a generate-and-verify strategy that constructs failure hypotheses from language model priors and visual evidence, discovers hypothesis slices using vision language models, and confirms them through statistical trend analysis. Evaluated on a new benchmark called GESD with expert-defined, spatially grounded slices, the method improved Precision@10 on detection to 0.73 from a baseline of 0.63. The work aims to produce interpretable error slices that guide targeted model improvements. Source