Reproduce “LLM-Guided ANN Index Optimization for Human-Object Interaction Retrieval” on OpenResearch

About this paper

Shahrzad Esmat, Chaunte W. Lacewell, Sameh Gobriel, Nilesh Jain, Ali Jannesari

An LLM-guided agent optimizes Approximate Nearest Neighbor (ANN) index and reranking parameters for multi-stage retrieval systems, achieving a 36.6x throughput increase over unoptimized baselines and a 33.3% performance improvement compared to leading traditional optimizers on a new Human-Object Interaction (HOI) retrieval benchmark.

Key insights

• LLM agents are effective at navigating and exploiting complex, coupled parameter dependencies in multi-stage retrieval systems, a task where traditional HPO methods often fail due to independence assumptions.
• The performance advantage of the LLM agent is proportional to the degree of parameter coupling in the search space, showing greater gains in highly coupled scenarios like HICO-DET compared to less coupled ones.
• Optimizing retrieval systems requires quality-constrained objectives (like SIEVE) to ensure proposed configurations are not only fast but also meet minimum acceptable quality thresholds, making them viable for real-world production.

Summary by alphaXiv.