Search papers, labs, and topics across Lattice.
1
0
3
Seemingly large gains from complex, learned retrieval models on multilingual and visual RAG benchmarks may simply be due to better document representation, as BM25 closes the gap when given the same high-quality input.
