Publications

We directly encode document screenshots into vectors using visual-LLM for semantic search, unifying multimodal retrieval approaches.

NEST enhances factuality and attribution of LLMs through nearest neighbor speculative decoding techniques.

We investigate how query expansion techniques can improve the generalization capabilities of strong cross-encoder ranking models.

CELI provides a simple yet effective approach to improve cross-encoder generalization across different domains.

DRAGON introduces diverse augmentation techniques to improve the generalization of dense retrieval systems across different domains.

CITADEL provides an efficient multi-vector retriever that is about 40x faster than ColBERT-v2 on GPUs through dynamic lexical routing.

SLIM reduces the latency and storage of ColBERT while being fully compatible with Pyserini (Lucene-based) indexing systems.

Aggretriever presents a simple approach to aggregate textual representations for building more robust dense passage retrieval systems.

We provide certified error control for candidate set pruning in two-stage relevance ranking systems.

We analyze encoder attributions in Dense Passage Retriever for open-domain question answering tasks.

We propose simple and effective unsupervised methods to compress dense vectors for passage retrieval.

We develop multi-task dense retrieval methods using model uncertainty fusion for open-domain question answering.

We reproduce DPR training and replicate retrieval results to provide insights into dense passage retrieval.