candidate generation and ranking

This work addresses the inefficiencies of conventional text chunking in standard retrieval-augmented generation (RAG) systems, which often introduces redundancy, leading to excessive storage costs and degraded retrieval performance. To mitigate this, the authors propose a lightweight pre-index filtering mechanism that integrates semantic similarity, topic coherence, and named entity recognition to selectively prune redundant text chunks prior to indexing. Evaluated through a token-level precision, recall, and intersection-over-union framework, the approach reduces vector index size by 25%–36% while preserving retrieval quality comparable to that of the original system, thereby significantly enhancing the overall efficiency of RAG pipelines.

Traditional retrieval-then-reranking pipelines suffer from two key limitations: dependency on the quality of initial retrieval and the high computational cost of large language model (LLM)-based rerankers. To address these, we propose Reranker-Guided Search (RGS), the first approach to explicitly incorporate reranker preferences into the retrieval process. RGS constructs a proximity graph over approximate nearest neighbors, then performs greedy path search guided jointly by embedding similarity and gradients of reranker scores—dynamically prioritizing high-potential documents. This breaks the rigid “retrieve-then-rerank” sequential paradigm and enables end-to-end optimization under a fixed reranking budget (100 documents). Evaluated on BRIGHT, FollowIR, and M-BEIR benchmarks, RGS improves Recall@100 by 3.5, 2.9, and 5.1 percentage points, respectively, significantly alleviating the precision-efficiency trade-off.

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 work addresses the challenge in LongEval-RAG tasks where responses must be strictly grounded in a given set of candidate documents. To this end, the authors propose a candidate-constrained retrieval-augmented generation (RAG) system that integrates rule-based chunking, query expansion, pseudo-relevance feedback, reciprocal rank fusion, MiniLM sentence-level reranking, and citation-aware evidence aggregation, complemented by deterministic provenance tracing and a neural sentence selection mechanism. Experimental results demonstrate that the proposed rule-MiniLM variant significantly outperforms baselines across multiple metrics—including BERTScore, retrieval precision, information point coverage, and human evaluation—thereby validating the effectiveness of combining rule-based chunking with neural sentence selection. The study further underscores the critical role of multi-metric evaluation in diagnosing and advancing RAG system performance.

In retrieval-augmented generation (RAG), a semantic gap exists between retriever rerankers and generative models in assessing document relevance. To bridge this gap, we propose RADIO, the first framework that jointly leverages large language model (LLM)-generated reasoning rationales and preference-aligned fine-tuning. Specifically, an LLM first extracts step-by-step reasoning traces required to answer the query; these rationales serve as supervision signals to rerank retrieved documents; finally, the reranker is further optimized via preference learning to align with generative behavior. This enables semantic-level coordination between reranking and generation. Experiments across three benchmark datasets and two task categories demonstrate that RADIO significantly improves answer quality and factual consistency over state-of-the-art baselines. The implementation is publicly available.

This work proposes a novel approach to large-scale retrieval that circumvents the prohibitive cost of full reranking by constructing query and item embeddings derived from the outputs of a reranker. Specifically, it leverages relevance scores assigned by a heavyweight reranker over a set of support items to generate lightweight embeddings, thereby enabling the reranking model to directly guide embedding learning—a capability demonstrated here for the first time. Under mild conditions, the method is theoretically shown to approximate arbitrarily complex similarity functions. Through systematic investigation of support item selection strategies and integration with approximate nearest neighbor search, the approach significantly improves candidate set quality across multiple academic and industrial datasets while maintaining computational efficiency.

This work addresses the limitations of large language models in million-scale candidate ranking, where performance is constrained by context length and computational cost. The authors propose LRanker, a novel framework that integrates K-means clustering with graph partitioning to construct candidate subsets and introduces a multi-query embedding generation and ensemble mechanism to enhance the model’s perception capacity at inference time. By modeling global information and aggregating embeddings through ensemble strategies, LRanker significantly improves ranking accuracy and robustness. Evaluated on the RBench benchmark, the method achieves over 30% performance gains in small-scale settings, 3–9% improvements in mean reciprocal rank (MRR) for million-scale scenarios, and maintains stable performance with 20–30% gains even in ultra-large-scale setups involving more than 6.8 million candidates, demonstrating its effectiveness and strong scalability.

This work addresses the limitations of current large language model (LLM)-based rerankers, which suffer from high computational overhead and context-length constraints, while conventional truncation strategies rely on static heuristics that lack dynamic awareness of query relevance. The authors propose a novel approach that leverages an LLM to generate a semantic reference document, serving as a dynamic boundary between relevant and non-relevant documents to guide list truncation. This is combined with either non-overlapping or adaptively stepped overlapping window mechanisms to enable efficient list-wise reranking. Notably, this is the first method to employ LLM-generated reference documents for dynamic truncation, overcoming the constraints of fixed hyperparameters and topic-agnostic heuristics. Evaluated on the TREC Deep Learning benchmark, the approach significantly outperforms existing truncation strategies, achieving up to 66% speedup in both in-domain and out-of-domain settings.

This work addresses the scarcity of scalable, low-cost, and controllable test collections for information retrieval systems, particularly in scenarios involving sensitive data or early-stage system design where human-annotated relevance judgments are unavailable. To this end, the authors propose SPECTRA, a novel framework that decouples latent topic structure, surface text generation, metadata control, query intent modeling, and deterministic relevance oracles, enabling fine-grained control over long-tail term distributions, distractor proportions, and relevance labels in synthetic corpora. Using a single-threaded Python prototype, SPECTRA efficiently generates a reproducible test collection comprising 60,000 documents and 9.61 million tokens at a rate of 12,000–14,000 documents per second. The framework demonstrates strong diagnostic utility: as distractor prevalence increases to 36%, BM25’s nDCG@10 drops from 1.00 to 0.43, effectively supporting stress-testing of retrieval systems.

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.