ResRank: Unifying Retrieval and Listwise Reranking via End-to-End Joint Training with Residual Passage Compression

Modern information retrieval systems typically employ a multi-stage pipeline where a lightweight first-stage retriever rapidly recalls candidate passages from a large corpus, followed by a more sophisticated reranker that refines the ranking order. With the advent of large language models, the reranking stage has witnessed remarkable progress: LLM-based listwise rerankers accept a query together with multiple candidate passages and directly output a permutation, substantially outperforming traditional cross-encoder approaches. However, deploying LLM-based listwise rerankers at scale faces two fundamental challenges: concatenating the full text of dozens or hundreds of candidate passages creates extremely long input sequences, leading to the 'lost in the middle' phenomenon, which directly undermines ranking quality. Even when the input length is manageable, the autoregressive decoding process—generating passage identifiers token by token—adds considerable overhead, especially when ranking long candidate lists.

Traditional large language models face significant efficiency bottlenecks when processing long input sequences. As input length grows, inference latency scales super-linearly, making it impractical for industrial deployment. Additionally, long input sequences lead to the 'lost in the middle' phenomenon, where information buried in the middle of long contexts is disproportionately neglected, directly undermining ranking quality. While sliding window strategies can partially alleviate this issue, the resulting multi-pass inference multiplies latency by a factor proportional to the number of windows, still failing to meet the demands of real-time applications.

ResRank's core innovations include:

1. Residual Passage Compression: Inspired by multimodal LLMs, ResRank employs an Encoder-LLM to compress each candidate passage into a single dense embedding, which is directly fed into the Reranker-LLM's input space. A residual connection combines the original encoder embedding with the contextualized hidden state produced by the reranker, reducing learning difficulty and preserving passage-level information.

2. Similarity-Based Scoring: Replacing costly token-by-token autoregressive decoding, ResRank adopts a retrieval-inspired scoring mechanism: the reranker's hidden state at the end-of-sequence position, enriched by cross-passage contextual attention, serves as a global aggregation embedding, which is directly compared with each passage's fused representation via a single-step cosine similarity computation.

3. Dual-Stage, Multi-Task, End-to-End Joint Training: ResRank trains the Encoder-LLM and Reranker-LLM in an end-to-end manner through a two-stage supervised fine-tuning process. Multi-task learning is employed to simultaneously optimize ranking and retrieval objectives, ensuring that the encoder preserves its retrieval capability while adapting to the reranker's requirements.

ResRank's methodology details are as follows:

• �� Passage Compression: Each candidate passage is compressed into a single embedding by the Encoder-LLM, reducing input length.
• �� Residual Connection: Combines the encoder embedding with the reranker's contextualized hidden state to form the fused passage embedding, preserving passage-level information.
• �� Similarity-Based Scoring: Uses cosine similarity to compute the relevance score between the global aggregation embedding and each fused passage embedding, eliminating autoregressive decoding.
• �� Dual-Stage Training: The first stage establishes coarse-grained alignment, while the second stage provides fine-grained refinement, ensuring the encoder's retrieval capability is preserved throughout the entire training process.

The experimental design includes evaluations on TREC Deep Learning 2019 and 2020 test sets and eight BEIR benchmark datasets. All models rerank the top-100 passages retrieved by BM25, using nDCG@10 as the primary evaluation metric. Training data is divided into two stages: the first stage trains on 232,419 samples, and the second stage trains on approximately 87,000 high-quality samples. Baseline models include a comprehensive suite of reranking models, covering supervised models, unsupervised LLM-based models, and distillation-trained LLM models.

Experimental results show that ResRank in single-pass mode surpasses RankMistral, ListT5-3B, and PE-Rank by over 5 absolute points in average nDCG@10 on BEIR datasets, particularly excelling on Signal and News datasets. On TREC Deep Learning 2019 and 2020 benchmarks, ResRank outperforms PE-Rank in single-pass mode and exceeds most distilled LLM rerankers. Ablation studies confirm that residual connections, dual-stage training, end-to-end optimization, and multi-task learning are indispensable to the final performance.

ResRank's application scenarios include:

• �� Real-Time Information Retrieval: ResRank significantly enhances ranking efficiency and effectiveness in scenarios requiring efficient processing of large candidate passages, suitable for search engines and recommendation systems.
• �� Industrial Search Engines: By reducing input length and eliminating autoregressive decoding, ResRank offers new possibilities for real-time information retrieval in industrial applications.
• �� Natural Language Processing Tasks: ResRank's compression algorithm and scoring mechanism can be applied to other NLP tasks requiring efficient processing of long texts, such as text summarization and question answering systems.

ResRank may still experience information loss when processing extremely long texts, particularly during the compression phase, where some details might be overlooked. Additionally, although ResRank performs well in multi-task learning, its training process requires substantial computational resources, which may not be suitable for resource-constrained environments. In certain specific domains or datasets, ResRank's performance might not match that of models optimized specifically for those areas. Future research could focus on further optimizing ResRank's compression algorithm to reduce information loss and exploring efficient training methods for ResRank in resource-limited environments.