Breaking01 / AI
Place search accuracy falls when source views change
AcrossPlace researchers report lower place-search accuracy when retrieval images differ from those used to generate descriptions.

AcrossPlace researchers reported in a preprint published on 13 Nov 2026, not yet peer-reviewed, that separating description-source images from retrieval targets reduced place-search accuracy to 12.31% (three training seeds, independent target images, unseen places within represented cities); the abstract reports no independent replication results.
The paper builds on GSV-Cities, separating the images used to generate descriptions from those searched for matching places. Searches from text to images and from images to text use the same criterion for deciding whether a result depicts the correct place.
For SigLIP2 trained with place information, the researchers report top-ranked text-to-image retrieval accuracy of 19.69% (three training seeds, source images, unseen places within represented cities). Accuracy falls to 12.31% (three training seeds, independent target images, the same geographic evaluation). This measure records whether the highest-ranked result matches the requested place.
The researchers also report cases where descriptions passed automated checks despite discrepancies with factual evidence. Their description audit indicates that factual inconsistencies remain after automated filtering.
The paper argues that earlier aggregate retrieval scores mix together changes in viewpoint, competition among candidate results and linguistic clues. AcrossPlace compares paired observations and controls the retrieval candidates to distinguish these effects.
The abstract does not disclose the rate of factual discrepancies found in the description audit.