MICE: Minimal Interaction Cross-Encoders for efficient Re-ranking
Researchers' new architecture bridges cross-encoder quality with late-interaction speed for search.
Deep Dive
Researchers from multiple institutions developed MICE (Minimal Interaction Cross-Encoders), a new AI architecture for document re-ranking. By strategically removing unnecessary attention interactions from standard cross-encoders, MICE achieves 4x lower inference latency while maintaining most in-domain effectiveness. It matches the speed of late-interaction models like ColBERT and shows superior generalization on out-of-domain datasets, making high-quality re-ranking practical for production systems.
Why It Matters
Enables search engines and RAG systems to use state-of-the-art ranking without prohibitive computational costs.