🧑‍🔬 Vincent Zhao

Vincent Zhao works on neural retrieval and on making the output of large language models more trustworthy. Their papers here ask how dense retrievers can carry over to new tasks and domains, and how a model's generated text can be tied back to evidence found on the web. The methods centre on scaling pretrained encoders, contrastive training, and few-shot prompting of large models, evaluated on standard retrieval and attribution benchmarks. The work sits mainly in Information retrieval & search, with a strand in Language models & pretraining.

Two of the most cited papers deal with dense retrieval. Large Dual Encoders Are Generalizable Retrievers (2021) tests whether dual encoders transfer poorly because a fixed-size dot product is too weak. It grows the T5-based encoder from Base to XXL while leaving the embedding size unchanged. The resulting GTR models are first pretrained contrastively on a large set of web-mined question-answer pairs and then fine-tuned. Promptagator (2022) questions the idea that one notion of relevance suits every retrieval task. It sets up a few-shot version of BEIR, where each task comes with only a short description and a handful of labelled query-document pairs.

The most recent work turns from retrieval itself to attribution for generated text. RARR: Researching and Revising What Language Models Say (2023), also the most cited of the papers here, introduces the task of editing for attribution. A system takes any model's output, uses web search to gather evidence through generated questions, and then makes minimal edits wherever the text and the evidence disagree. It needs no changes to the original model. Zhao's current affiliations are Google and The University of Texas at Austin. The most frequent co-author is Zhuyun Dai, and other collaborators include Jianmo Ni, Ming-Wei Chang, Kelvin Guu and Keith Hall.

Vincent Zhao has 3 papers in our library, in 2021–2023, most of them on Information retrieval & search, Language models & pretraining; often with Zhuyun Dai, Ji Ma, Jianmo Ni.