Optimizing Listwise Reranking with ModernBERT: A Data-Centric Approach with a Novel Graded Corpus

This thesis investigates how ModernBERT can be used as an effective cross-encoder reranker within contemporary information retrieval systems. To enable this, a graded-relevance dataset was constructed, offering richer and more expressive supervision than traditional binary relevance labels. This dataset allows the model to learn subtle distinctions in relevance and better capture the nuanced relationships between queries and documents. The reranker is evaluated in a zero-shot setting on a broad selection of BEIR datasets spanning multiple domains. Despite never being fine-tuned on these tasks, ModernBERT consistently strengthens the final rankings produced by the retrieval pipeline, demonstrating an ability to transfer its learned ranking behaviour beyond the supervised training distribution. The improvements are especially clear on datasets whose structure and content align naturally with the type of signals present in the constructed training data. Overall, the thesis shows that ModernBERT is a capable and adaptable reranking model, benefiting significantly from graded supervision and generalising well across diverse retrieval scenarios. These findings highlight the value of high-quality relevance annotations and point toward future work on expanding supervision sources and refining reranking architectures for even stronger performance.