LightOn AI built a 2-billion-parameter multimodal reranker using a generative listwise approach that handles both text passages and document page images with a single adapter. The model scored 62.66 NDCG@10 on the ViDoRe V3 benchmark, outperforming comparable 2B-class open rerankers through cross-document attention mechanisms that compare candidates against each other. The team found that standard text-reranker speedup techniques such as tournament scheduling failed for the vision variant, because removing head-to-head comparison between candidates undermines the very mechanism that gives the model its accuracy.
