Search papers, labs, and topics across Lattice.
This paper introduces RSRank, a novel reranking approach that leverages representational shifts (RS) in a query's internal state to assess document relevance, moving beyond traditional heuristic thresholds. By establishing a correlation between RS induced by candidate documents and an oracle document set, the authors create a robust relevance indicator. The proposed lightweight training framework not only reduces reliance on heuristic tuning but also demonstrates significant performance improvements over state-of-the-art rerankers across various retrieval datasets.
RSRank reveals that representational shifts can serve as a powerful and principled signal for document relevance, outperforming existing heuristic-based methods.
As enterprises deploy RAG-based systems to provide grounded responses to user queries, reranking has become a critical component for the final filtering step that separates relevant from distracting or irrelevant documents. Existing rerankers often rely on heuristic thresholds to achieve optimal filtering. Moreover, for relevance scoring, state-of-the-art methods use a language model's logit signals, which are designed for next-token prediction, not for assessing relevance. To address these limitations, we identify a principled signal for relevance: the representational shift (RS) induced in a query's internal state when conditioned on a document. We observe that the alignment between (a) RS induced by a candidate document and (b) RS induced by an oracle document-set provides a robust indicator of relevance. Building on this insight, we introduce a lightweight training framework that learns projections mapping RS to calibrated relevance scores. Our training objectives naturally filter irrelevant content at a zero threshold, reducing dependence on heuristic tuning. Across diverse retrieval datasets, our method delivers gains over SOTA rerankers.
