Multimodal retrieval often relies on single-vector retrievers, which encode an entire input into a single global representation. According to a new arXiv preprint, this approach is efficient but loses important local detail.
The abstract explains that sequential token sequences are condensed into one vector, and while that helps performance, fine-grained evidence critical for retrieval is discarded. The paper's title indicates a method called SMARTer, which presumably counteracts this trade-off.
Because the abstract is truncated, the specific mechanism behind SMARTer is not described here. The preprint is listed as arXiv:2605.24938v2 and appears to be a replacement, though the version history is not detailed.