LightOn-rerank Targets Mixed-Corpus RAG With Multimodal Rerankers
LightOnIO has released LightOn-rerank, a family of multimodal rerankers designed for mixed-corpus RAG. The models can score both query–text-passage relevance and query–document-page-image relevance, making them notable for workflows where PDFs, scanned pages, and plain text coexist in the same knowledge base.
Key details
The series includes 6 models across three sizes: 0.8B, 2B, and 4B. Each size comes in both pointwise and listwise variants. Posts describe the 2B models as especially competitive among open rerankers of similar size; one cited result is that the pointwise 2B model reaches 59.8 on ViDoRe V3. The models use a single adapter to support both text and document-page inputs.
Training and deployment status
According to the shared write-ups, training used 213K training groups and only a single epoch. Commentators noted this leaves clear headroom versus approaches trained on millions of samples or distilled from larger teachers. The training data was English, but the models reportedly show some cross-lingual generalization. For deployment, current Sentence Transformers support does not yet handle multi-document listwise reranking, so the pointwise versions are easier to use immediately.
Ablations and limitations
The accompanying discussion also highlighted approaches that underperformed. Tournament scheduling reportedly reduced results by 27 NDCG, while first-token readout delivered only limited gains. That disclosure gives a clearer picture of both the current boundary of the method and where future optimization may focus.
2026-07-16 ~ 2026-07-17 · 10 related posts
- Collection of Multimodal Rerankers Released — CShorten30 · 2026-07-16
- LightOn Releases Multimodal Reranker — antoine_chaffin · 2026-07-16
- Update on Multimodal Reranker Research — antoine_chaffin · 2026-07-16
- LightOn Unveils Reranker Model Series — tomaarsen · 2026-07-17
- Pointwise Rerankers: Faster and Surprisingly Strong — tomaarsen · 2026-07-17
- Reranker Designs That Failed to Deliver — tomaarsen · 2026-07-17
- Room to Improve in Lightweight Reranker Training — tomaarsen · 2026-07-17
- Strong New RAG Solution for Mixed Corpora — tomaarsen · 2026-07-17
2 near-duplicate retellings: CShorten30 · tomaarsen