Reproduce “ELVA: Exploring Ranking-Driven Universal Multimodal Retrieval” 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
Yuhan Liu, Pei Fu, Hang Li, Yukun Qi, Chao Jiang, Jingwen Fu, Zhen Liu, Bin Qin and 3 more
Researchers from Xi'an Jiaotong University and Xiaomi Inc. developed ELVA, a ranking-driven framework that mitigates "grain blindness" in Universal Multimodal Retrieval (UMR) systems by leveraging Multimodal Large Language Models (MLLMs) and a novel reinforcement learning (RL) approach with rule-based verifiable rewards. ELVA achieved state-of-the-art performance on M-BEIR and showed a 13.1% improvement on the new MRBench, designed for multi-grain retrieval.
Key insights
- "Grain blindness" in MLLMs results from "Gradient Starvation" during contrastive learning, causing secondary semantic grains to lose discriminative signals.
- Rule-based, continuous reward functions in a reinforcement learning setup can effectively incentivize models to learn nuanced inter-negative hierarchies, improving fine-grained ranking without explicit labels.
- The RL stage acts as a modular "plug-and-play" enhancement that can improve the performance of existing multimodal retrievers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.