Zhuchenyang Liu

Stored patch vectors from visual document retrieval systems can be inverted to reconstruct original pages, revealing sensitive information and enabling source page identification with high accuracy.

Distilling multi-billion-parameter multi-vector retrievers no longer requires caching terabytes of document tokens: entropic optimal transport allows a lightweight 149M text model to capture 95% of VLM teacher performance at 26x query speed using queries alone.

A single-vector visual document retrieval system that is 15.6 times smaller and an order of magnitude faster than existing multi-vector models while retaining high accuracy.

Quoting evidence directly from documents boosts evidence recall by up to 39 points while halving hallucination rates, challenging the reliance on coordinate-based methods.

VLMs struggle to align assembly diagrams and videos because they occupy disjoint visual representation spaces, revealing a fundamental limitation in cross-modal understanding.

Shrinking a 2B vision-language retriever to a 70M text-only model achieves 95% of the original quality and outperforms a 2B baseline, while slashing query latency by 50x.

Ditch global embeddings for text-motion retrieval: this method uses joint-angle motion images and token-patch late interaction to achieve state-of-the-art accuracy and interpretability.