EEG-EditBench Diagnoses Visual Information in EEG-Retrieval Models
July 31, 2026
EEG-EditBench introduces a diagnostic benchmark using 2,137 controlled image edits to test if EEG-to-image retrieval models rely on fine-grained visual attributes. Results show that high retrieval performance does not guarantee a model can distinguish specific object identity or attribute changes.
HOW THIS AFFECTS YOU
●
researcherYou should use this benchmark to validate whether your neural decoding models are capturing semantic attributes or merely coarse scene context.