RankEvolve: Automating the Discovery of Retrieval Algorithms via LLM-Driven Evolution
An AI system evolved novel retrieval algorithms that outperform classic methods like BM25 on 12 IR benchmarks.
Deep Dive
Researchers from Santa Clara University and Meta introduced RankEvolve, an LLM-driven evolutionary system. Starting from seed algorithms like BM25, it uses mutation and selection to generate executable ranking code. The system discovered novel algorithms that showed promising transfer performance across 12 datasets from BEIR and BRIGHT, suggesting a path toward automating the discovery of foundational IR techniques.
Why It Matters
This could automate the design of core search infrastructure, moving beyond human-tuned algorithms to AI-optimized ones.