Please note: This PhD seminar will take place online.
Ronak Pradeep, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Jimmy Lin
Rerankers sit at the heart of modern information access systems, refining candidate documents from first-stage retrievers into well-ordered results that are often presented to users. Despite their critical role, the state-of-the-art in the field has recently been constrained by reliance on proprietary and resource-intensive models.
This seminar presents a progression of open-source reranking research culminating in RankZephyr, a 7B-parameter listwise reranker that matches or surpasses GPT-4-level effectiveness while remaining open and reproducible. We explore design insights from supervised cross-encoders to large-scale instruction-distilled listwise models, examining how prompt design, loss functions, and data curation impact generalization. Together, these results demonstrate that open rerankers can be both powerful and efficient, paving the way for accessible and highly-effective retrieval pipelines in the LLM era.