Search papers, labs, and topics across Lattice.
2
0
5
2
Flipping relevance labels via LLM-generated complementary instructions boosts instruction-following retrieval by 45%, proving that targeted data synthesis beats brute-force scaling.
Code-switching can degrade information retrieval performance by up to 27%, revealing a critical blind spot in current multilingual models.