RankEvolve: How AI Discovers Better Search Algorithms Through Evolution

📄 RankEvolve: Automating the Discovery of Retrieval Algorithms via LLM-Driven Evolution

🎯 What This Research Is About

Retrieval algorithms like BM25 and query likelihood with Dirichlet smoothing have been the backbone of information retrieval for years. While they remain strong and efficient, improvements have mostly relied on manual parameter tuning and human intuition. This paper introduces RankEvolve, a groundbreaking approach that uses large language models (LLMs) guided by evaluators and evolutionary search to automatically discover improved lexical retrieval algorithms.

RankEvolve is based on AlphaEvolve and represents candidate ranking algorithms as executable code. These algorithms are iteratively mutated, recombined, and selected based on their retrieval performance across 12 IR datasets from BEIR and BRIGHT benchmarks.

💡 Why This Matters

• Automated Algorithm Discovery: Instead of relying on manual engineering, RankEvolve demonstrates that LLMs can automatically evolve novel ranking algorithms that outperform traditional baselines
• Strong Generalization: The evolved algorithms show promising transfer capability to the full BEIR and BRIGHT benchmarks, as well as TREC DL 19 and 20 datasets
• Novel Approach: This work opens a new research direction where AI systems can autonomously discover and improve information retrieval algorithms through program evolution
• Practical Impact: The methods could lead to better search engines, recommendation systems, and information retrieval applications across industries

📖 Read Full Paper →
💻 View Code on GitHub →

---

Curated from Hugging Face daily papers by AMS IT Services AI Research Team