ResRank: A Unified Approach to Retrieval and Listwise Reranking

Recent advancements in large language models (LLMs) have positioned listwise reranking as a leading method for improving ranking performance in information retrieval systems. However, traditional approaches face significant hurdles, particularly the 'lost in the middle' issue, which diminishes ranking accuracy as input sizes increase, and the excessive latency associated with longer sequences, complicating real-world applications. ResRank offers a solution by integrating retrieval and reranking into a cohesive framework. Drawing inspiration from multimodal LLMs, ResRank compresses candidate passages into compact embeddings using an Encoder-LLM. These embeddings are then processed with the query by a Reranker-LLM for effective listwise ranking. To bridge the gap between the compressed representations and the ranking process, a residual connection is implemented, merging encoder outputs with the reranker’s contextual states. Additionally, the framework replaces traditional autoregressive decoding with a streamlined cosine-similarity scoring method, effectively removing generation delays. ResRank employs a dual-stage, multi-task joint training approach, aligning the learning objectives of both retrieval and reranking while simplifying the training process. Comprehensive evaluations on TREC Deep Learning and various BEIR datasets indicate that ResRank not only matches but often surpasses existing methods, achieving a superior balance of ranking effectiveness and operational efficiency.