Systematic failures of imaginative and prescient fashions on semantically coherent subsets, generally known as error slices, reveal limitations in robustness and analysis. Current slice discovery approaches largely mannequin slices as clusters in illustration house or mixtures of predefined attributes. Whereas efficient for image-level classification, such formulations are inadequate for instance-level duties reminiscent of object detection and segmentation, the place failures usually come up from contextual relational and spatially grounded visible patterns. We suggest GH-ESD (Grounded Speculation-Pushed Error Slice Discovery), a generate and confirm framework that reformulates slice discovery as grounded speculation era and statistical verification. GH-ESD constructs relational failure hypotheses utilizing LLM priors and grounded visible proof, discovers speculation slices on the occasion degree by way of Imaginative and prescient Language Fashions, and verifies them by way of statistical pattern evaluation over instance-level errors. We additionally introduce GESD (Grounded Error Slice Dataset), a brand new benchmark for instance-level error slice discovery, offering expert-defined and spatially grounded slices derived from detection and segmentation failures. Intensive experiments display that GH-ESD constantly outperforms baselines, enhancing Precision@10 by 0.10 (0.73 vs. 0.63) on the GESD benchmark for detection duties, whereas additionally supporting segmentation eventualities. GH-ESD identifies interpretable slices that facilitate actionable mannequin enhancements.
* Equal contribution.
