The Evolution of Reranking Models in Information Retrieval: From Heuristic Methods to Large Language Models

This version of the article has been accepted for publication in Springer CCIS, volume 2775, after peer review (when ap- plicable) but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at springnature. Use of this Accepted Version is subject to the publisher’s Ac- cepted Manuscript terms of use https://www.springernature.com/gp/open-research/policies/accepted-manuscriptterms. The Evolution of Reranking Models in Information Retrieval: From Heuristic Methods to Large Language Models Tejul Pandit1[0009-0006-4376-1063], Sakshi Mahendru1∗, Meet Raval2∗, and Dhvani Upadhyay3 1 Palo Alto Networks, USA 2 University of Southern California, USA 3 Dhirubhai Ambani University, India tejulpandit96@gmail.com Abstract. Reranking is a critical stage in contemporary information re- trieval (IR) systems, improving the relevance of the user-presented final results by honing initial candidate sets. This paper is a thorough guide to examine the changing reranker landscape and offer a clear view of the advancements made in reranking methods. We present a comprehensive survey of reranking models employed in IR, particularly within mod- ern Retrieval Augmented Generation (RAG) pipelines, where retrieved documents notably influence output quality. We embark on a chronological journey through the historical trajectory of reranking techniques, starting with foundational approaches, before exploring the wide range of sophisticated neural network architectures such as cross-encoders, sequence-generation models like T5, and Graph Neural Networks (GNNs) utilized for structural information. Recognizing the computational cost of advancing neural rerankers, we analyze tech- niques for enhancing efficiency, notably knowledge distillation for creat- ing competitive, lighter alternatives. Furthermore, we map the emerging territory of integrating Large Language Models (LLMs) in reranking, ex- amining novel prompting strategies and fine-tuning tactics. This survey seeks to elucidate the fundamental ideas, relative effectiveness, compu- tational features, and real-world trade-offs of various reranking strate- gies. The survey provides a structured synthesis of the diverse rerank- ing paradigms, highlighting their underlying principles and comparative strengths and weaknesses. Keywords: Rerankers, Information Retrieval (IR), Retrieval Augmented Generation (RAG), Learning-to-rank, Neural rerankers, cross-encoders, T5, Graph Neural Networks (GNN), knowledge distillation, Large Lan- guage Models (LLM) * Equal contribution arXiv:2512.16236v1 [cs.IR] 18 Dec 2025 This version of the article has been accepted for publication in Springer CCIS, volume 2775, after peer review (when ap- plicable) but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at springnature. Use of this Accepted Version is subject to the publisher’s Ac- cepted Manuscript terms of use https://www.springernature.com/gp/open-research/policies/accepted-manuscriptterms. 2 T. Pandit et al. 1 Introduction Information retrieval (IR) systems are essential in today’s digital world, with their ability to power anything from online search and recommender systems to sophisticated question-answering and knowledge management platforms. A crit- ical component within these systems is the reranking stage. Reranking carefully reorders these candidates to show the user the most pertinent results after an initial, often rapid, retrieval phase that produces an extensive collection of can- didate documents. This dramatically improves the overall quality and efficacy of the IR system[1]. Notably, the significant development in Large Language Models (LLMs) has subsequently driven substantial improvements and interest in Re- trieval Augmented Generation (RAG) pipelines, where the precision of retrieved context, refined by reranking, is crucial for generating accurate and relevant outputs[2]. Based on [3], we represent the typical placement of the reranking component within a RAG pipeline in Figure 1. Fig. 1. RAG approach highlighting a post-retrieval step of Reranking documents. Reranking methods have evolved significantly, from heuristic scoring and traditional learning-to-rank models to deep learning and, more recently, LLMs, enabling richer semantic understanding and inter-document relationships. This survey provides a comprehensive exploration of the evolving field of reranking models in IR. We trace the development trajectory from classical learning-to-rank methods in Section 3, followed by Deep learning rerankers in Section 4. We also cover methods aimed at improving the efficiency of reranking models, with a particular focus on knowledge distillation techniques in Section 5. Finally, we examine the cutting-edge integration of LLM-based rerankers in Section 6. This study aims to provide scholars and practitioners with a

This content is AI-processed based on open access ArXiv data.