BCIwiki (bciwiki.com) — A candidate-score fusion framework called CORTIVA lifted Top-1 accuracy in EEG-to-image retrieval to 73.5%, according to a preprint posted to arXiv on August 2, 2026. The authors say most existing systems collapse heterogeneous visual supervision into a single embedding before ranking, imposing one similarity geometry on every candidate order. CORTIVA instead aligns three decoding routes to heterogeneous visual targets, scores the same indexed candidates independently, and combines only their temperature-scaled score vectors before ranking, preserving complementary evidence.
On the 200-way THINGS-EEG2 benchmark, the framework reaches 73.5% Top-1 and 95.3% Top-5 across ten participants. With a modality-specific neural encoder, the same fusion principle reaches 42.4% Top-1 on THINGS-MEG. The authors add that matched route-removal retraining and four weight controls indicate the gain arises from integrating complementary route scores and persists with uniform weighting, without requiring a specialized weighting rule. The work is a preprint and has not yet been peer reviewed.