apply neural reranking

This work addresses the challenge of ineffective reranking in dense retrieval systems under zero-shot scenarios, where supervised signals are absent. The authors propose DART, a novel method that performs lightweight adaptive training at test time to refine reranking. Specifically, DART generates pseudo-labels from top- and bottom-ranked documents in the initial retrieval results and fine-tunes the bilinear scoring matrix via a small number of gradient updates, guided by a confidence-weighted margin loss and a cross-query momentum buffering mechanism. Requiring no additional annotations, DART achieves an average relative improvement of 2.1% in NDCG@10 across six BEIR benchmarks, with less than 10ms added latency per query.

Existing neural re-rankers achieve strong performance but suffer from high query-time computational overhead, poor generalization to complex queries, and multilingual support requiring task-specific fine-tuning. This paper proposes Rank-K—the first listwise re-ranker enabling test-time reasoning—where computation is dynamically allocated per query to achieve adaptive refinement. Its core innovation lies in natively integrating reasoning-capable large language models into a listwise ranking framework, coupled with multilingual unified representation learning and contrastive alignment, eliminating the need for fine-tuning to achieve cross-lingual re-ranking. Experiments demonstrate that, applied to BM25 initial rankings, Rank-K outperforms the state-of-the-art RankZephyr by 23% in NDCG@10; when initialized from the strong retriever SPLADE-v3, it yields a 19% gain. Crucially, Rank-K maintains monolingual effectiveness while achieving robust multilingual transfer—without any language-specific adaptation.

This study identifies a performance breakpoint and semantic failure in cross-encoder re-rankers (e.g., ColBERTv2, RankT5) for large-scale document re-ranking: retrieval quality degrades significantly when the candidate set exceeds ~1,000 documents—MRR@10 drops by 12.7% on average, and 38% of top-scoring results exhibit neither lexical overlap nor semantic similarity with the query. Through systematic ablation and scaling experiments, augmented with semantic similarity and lexical matching analyses, we empirically challenge the widely held assumption that re-rankers universally outperform first-stage retrievers. Our key contributions are: (1) establishing the effective scale boundary for cross-encoder re-rankers; (2) revealing their propensity for relevance misjudgment under ultra-large candidate lists; and (3) providing theoretical grounding and practical guidance—along with critical deployment warnings—for integrating re-ranking modules into large-scale retrieval systems.

This study addresses the challenge of medical procedure reranking caused by the lexical gap between patient queries and clinical terminology. For the first time in a real-world health insurance setting, it systematically compares the performance and efficiency of lightweight cross-encoders against large language model (LLM)-based instruction rerankers. The authors fine-tune domain-specific models such as MedCPT and MiniLM-L12 using listwise ranking losses like ListNet and introduce a GPT-4–driven agent-based prompt optimization pipeline to enhance the Qwen3-Reranker-4B. Experimental results demonstrate that a compact cross-encoder with only 109 million parameters outperforms the 4-billion-parameter LLM by 2.6 points in NDCG@3 and by 13.3 points in Spearman correlation, while using merely 1/37th of the parameters—highlighting the efficacy and scalability of small models for specialized domain tasks.

To address high latency and computational overhead in LLM-based pairwise re-ranking for real-time retrieval-augmented generation, this paper proposes a systematic optimization framework tailored for real-time deployment. Methodologically, it replaces large LLMs with lightweight variants, strictly constrains the candidate set size for re-ranking, applies INT4 quantization, designs a unidirectional sequential inference architecture to mitigate positional bias, and caps output length. Unlike prior approaches, this work achieves the first end-to-end real-time LLM-driven pairwise re-ranking (<0.4 s/query), reducing latency by 166× (from 61.36 s to 0.37 s) while preserving Recall@k nearly losslessly. Extensive experiments reveal that several previously overlooked yet critical design choices—particularly those governing inference architecture, quantization, and candidate pruning—exert decisive influence on the efficiency–effectiveness trade-off. The framework significantly enhances the practical deployability of LLM-based re-ranking in production environments.

This work addresses the sensitivity of large language models (LLMs) to the input order of candidate items in recommendation reranking, which leads to inconsistent ranking results. To resolve this issue, the authors propose InvariRank, a framework that enforces permutation invariance at the architectural level by employing a structured attention mask to block cross-attention among candidates and integrating Rotary Position Embeddings (RoPE) with shared positional encoding to eliminate positional bias. This design enables listwise reranking in a single forward pass while guaranteeing positional invariance—without relying on multi-permutation training strategies. InvariRank achieves competitive performance across multiple recommendation benchmarks and significantly enhances ranking stability under varying input permutations.

This work addresses the inefficiency of traditional retrieval systems that uniformly apply high-cost reranking models to all queries, incurring unnecessary latency and computational overhead for simple queries. The authors propose a utility-based adaptive reranking framework that dynamically selects reranking strategies according to query complexity, enabling cost-aware query routing. A novel utility function is introduced to guide routing decisions, and the approach leverages BM25 for sparse retrieval, MiniLM-L6-v2 for lightweight dense reranking, and BGE-v2-m3 for heavyweight neural reranking. A trained routing classifier enables multi-tier reranking strategy selection. Compared to applying the full BGE model universally, the proposed method reduces median latency by 1.15× to 53× and average latency by 1.11× to 5.22×, with nDCG@10 varying between –17.5% and +4.0%, demonstrating competitive effectiveness across multiple datasets.

This work addresses the limited interpretability and inefficient context utilization in traditional rerankers, which output only scalar relevance scores without refined evidential support. To overcome this, the authors propose Prism-Reranker, a family of multi-scale models built upon Qwen3.5 that jointly models contribution statement generation and evidence passage distillation, yielding both relevance judgments and structured explanations. The approach leverages LLM-as-Judge data curation, hybrid training on real and synthetic queries, keyword-based query rewriting, and a combination of pointwise distillation with supervised fine-tuning to substantially enhance generalization. Evaluated on BEIR-QA subsets, the enhanced Qwen3-Reranker-4B achieves an average NDCG@10 improvement of 1.54, with both contribution clarity and evidence quality validated by LLM-based assessment. The models and training pipeline are publicly released.

This work addresses the limitation of traditional retrieve-then-rerank pipelines, whose effectiveness is constrained by the recall ceiling of the initial retriever and struggles with queries requiring complex reasoning. To overcome this, the authors propose Graph-based Adaptive Reranking (GAR), a novel approach that introduces an iterative exploration mechanism leveraging a corpus graph during reranking, thereby enhancing reasoning capabilities without modifying the initial retriever. GAR is the first method tailored for reasoning-intensive retrieval scenarios, demonstrating strong generalization across diverse reranking models. Evaluated on the BRIGHT benchmark, it achieves substantial performance gains while incurring minimal computational overhead, marking a significant step toward practical deployment of such systems.