Reproduce “ReinPool: Reinforcement Learning Pooling Multi-Vector Embeddings for Retrieval System” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sungguk Cha, DongWook Kim, Mintae Kim, Youngsub Han, Byoung-Ki Jeon, Sangyeob Lee

Researchers at LG Uplus developed ReinPool, a reinforcement learning framework that compresses multi-vector embeddings for retrieval systems. This method achieves over 1000x reduction in storage cost while recovering 76-81% of the original retrieval performance, making high-precision multi-vector models deployable at scale.

Key insights

  • Dynamic, learned selection of multi-vector embeddings via reinforcement learning is vastly superior to static pooling methods for preserving retrieval effectiveness under aggressive compression.
  • An inverse retrieval objective, combined with synthetic query generation and NDCG rewards, provides an effective self-supervised training signal for optimizing vector selection for retrieval without manual labels.
  • Aggressive compression of multi-vector embeddings into single, retrieval-optimized vectors is achievable, bridging the gap between high-precision late-interaction models and the efficiency required for large-scale, real-world deployment.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.