Why tiny late interaction models win

A 150M parameter late interaction model from LightOn outperforms models up to 8B parameters on the BrowseComp benchmark, which tests complex research-style queries. Late interaction models win by learning query and document representations separately, then computing token-level interactions only at the final step (MaxSim). This avoids the expensive all-to-all token interactions of traditional cross-encoders, allowing much smaller models to be both accurate and fast. The approach is especially valuable for AI agents that need efficient semantic reranking to find relevant information quickly without multiple search iterations.