CS Paper Radar

Optimal Fusion Weights Form Intervals, Not Points: A Latency-Aware Study of Query-Adaptive WRRF

Iris Ma, Zahra Shekarchi, Stuart Mcilroy, Mahtab Farrokh, Fattane Zarrinkalam, Joel Stremmel

cs.IR ·submitted 2026-10-17 00:18 UTC · v1
code/data

Abstract

Reciprocal Rank Fusion (RRF) is widely used to combine sparse and dense retrievers in hybrid information retrieval, but a critical limitation is that it assigns a fixed, uniform weight to each retriever regardless of the query. Query-aware methods treat the per-query optimal weight as a single point, an assumption whose empirical validity has not been examined. We present an empirical study of eight dynamic fusion weight prediction strategies across four datasets and four sparse--dense retriever pairs. When we compare these approaches to static fusion and a per-query oracle that bounds achievable performance, several findings emerge. First, the per-query optimal fusion weight is typically an interval rather than a point. Queries can have multiple optimal intervals, and these can take on a variety of shapes. Second, 74 of 128 dynamic fusion methods tested were significantly better than standard RRF after false discovery rate correction. Third, dynamic fusion methods exhibit differing quality and latency characteristics that should be considered when choosing an appropriate fusion method. To address this last finding, we present a tiered decision framework that maps practitioner constraints, dataset properties, label availability, hardware, and latency budget to an appropriate method class. We release the code for our experiments at https://github.com/thomsonreuters/query-based-rrf.

Closest papers

Nearest neighbours by meaning across the whole index.

The Fusion Tax: A Diagnostic Analysis of When Hybrid Retrieval Hurts

cs.IR cs.CL ·2026-10-22 ·Aaditya Chauhan, Arun Agarwal
publishedhot direction similarity 0.72
Abstract

Hybrid retrieval, which fuses a sparse lexical retriever with a dense neural one, is widely recommended in the software documentation surveyed in September 2026, usually with reciprocal rank fusion (RRF), on the strength of aggregate gains. We ask what fusion costs on individual queries, treating "do not fuse" as a first-class outcome, in the setting in which hybrid retrieval is deployed: a standard Anserini BM25, five dense encoders spanning 22M-596M parameters and three model generations, a learned-sparse (SPLADE) alternative, and a cross-encoder second stage, on seven BEIR datasets and three fusion functions. Even the best method, dense-weighted score interpolation, ranks relevant documents below the better of its two components on 13-48% of queries and below the deployable dense-only retriever on 8-21% (14-48% and 11-25% at a weight selected under leave-one-corpus-out validation), while beating dense-only in aggregate on every corpus with the primary encoder; the downside is significant under loss-averse risk-sensitive evaluation on six of the seven corpora. We call this loss the fusion tax. It is mostly a ranking effect that preserves recall in aggregate, it is most likely on the queries that only one retriever covers, and it persists across the encoder ladder and the SPLADE substitution. RRF carries the higher tax wherever the dense retriever is clearly stronger and only beats dense-only where the two components are balanced. Fusion failure is weakly predictable within a corpus but transfers unreliably across corpora, because the features that predict it change sign with the sparse-dense balance. A cross-encoder recovers part of the tax only where it is itself reliable. We distill these findings into guidance that keys the fusion decision to the measured sparse-dense balance of the target corpus.

0.38

Top-K Is Not a Budget for Hybrid Retrieval

cs.IR ·2026-09-14 ·Chunran Zhang
code/data similarity 0.70
Abstract · 5 pages, 2 figures, 3 tables. Code: https://github.com/ln-one/top-k-is-not-a-budget

Modern hybrid retrieval for RAG typically fuses the Top-$L$ results from dense and sparse retrievers, but a fixed truncation depth may not transfer across changing queries and corpora. Exact fusion removes the dependence on a fixed depth, yet completing a specified Top-$K$ still incurs variable access costs. We present DiBud, which takes an access budget directly as input and incrementally certifies and returns an exact prefix of the RRF ranking over the full lists. Selective access increases certified output within the budget, while budgeted stopping bounds accesses per request. Experiments on five query sets reveal long-tailed costs for completing exact Top-20. At a budget of 2048 accesses, DiBud increases mean certified output within the first 100 positions by 7.86% over balanced access. After budget calibration for 95% quality retention, held-out queries retain 95.05%--97.68% of mean nDCG@20 while using 65.92%--99.53% fewer accesses than completing exact Top-20.

0.43

XViewRank: Overlap-Conditioned Score Fusion for LLM-Augmented Sparse Retrieval

cs.IR ·2026-10-14 ·Xinxuan Lv, Minghan Li, Guodong Zhou
similarity 0.67
Abstract · 20 pages, 2 figures, 12 tables

Large language models can reduce vocabulary mismatch in sparse retrieval by generating alternative views of a query. Once these views are searched independently, however, their ranked lists may be complementary, redundant, or affected by query drift, making fixed fusion suboptimal. We present XViewRank, a lightweight retrieval-side method for conditionally fusing BM25 results from an original query, a hypothesis view, and a step-back view. For each pooled candidate, XViewRank retains three normalized BM25 scores and three query-level overlaps between the retrieved result sets. A small nonlinear scorer uses the overlaps as context that changes how the view-specific scores are interpreted, without jointly encoding query and passage text. Under a matched-input protocol, XViewRank improves nDCG@10 and MAP numerically over reciprocal rank fusion on both TREC DL 2019 and 2020, with a statistically significant MAP gain on DL19. Additional analyses show that the clearest benefit occurs when the view rankings have low agreement, while the evidence on DL20 is weaker. The results support overlap-conditioned score fusion as a compact retrieval-side alternative to fixed fusion, while also delimiting where its advantage is empirically established.

0.15

Parameterized Dense-Sparse Fusion for Hybrid Retrieval: Tuning a Rank-Score Mix on BEIR SciFact with Qdrant

cs.IR cs.AI ·2026-09-19 ·Satyanarayan Pati, Srikanth Patil
similarity 0.66
Abstract

We study a parameterized hybrid ranker that fuses a dense embedding list and a sparse lexical list. The method has a small, explicit parameter vector: a dense prior $α\in [0,1]$, a score-versus-rank mix $λ\in [0,1]$, an RRF smoothing parameter $κ> 0$, optional list-geometry coefficients that move $α$ per query, and a router margin $τ$ that can turn sparse search off. We grid-search those ranges on SciFact train (809 queries) and freeze the chosen values on SciFact test (300). The tuned rank-score mix ($α= 0.8$, $λ= 0.75$, $κ= 20$) reaches 0.753 nDCG@10 and 0.889 recall@10, outperforming dense BGE (0.742 / 0.871) and equal-weight RRF (0.707 nDCG@10) on that test split. A list-conditioned $α$ adds +0.0006 nDCG; a sparse-off router is rejected by the same train split (any $τ$ that skipped approximately 50% of queries lost nDCG). These coefficients are dataset-specific. Equal RRF with the same models does not beat dense on a nine-zip BEIR macro-average (0.479 vs. 0.519 nDCG@10). Repeating the same train-then-freeze sweep independently on all 20 indexed units beats equal RRF on 20/20 and dense on 16/20 (unit-mean nDCG@10 0.467 vs. 0.462 dense vs. 0.420 RRF). Other corpora should reuse the ranges, not a copy of the SciFact point.

0.21

RAGCompass: Generalizable Retriever Routing via Reranking

cs.CL cs.IR ·2026-11-03 ·Valentin Six, Mufei Li, Dongqi Fu, Yinglong Xia, Hong Li +1 more
similarity 0.63
Abstract

Different retrievers exist for retrieval-augmented generation, yet none performs best for all queries. Therefore, the key question is: given a query and retrieved passages from different retrievers, which one is most suited for that query? Existing routers usually make predictions from query-specific features and a fixed pool of retrievers; such routers have no output for an unseen retriever and are incapable by design on a new corpus. To fill these gaps, RAGCompass reads the retrieved passage sets themselves: passages from an unfamiliar retriever or corpus stay directly comparable to other passages, so routing stays well designed where a query-only map would fail. We contrastively fine-tune a reranker against the answer each evidence set produces, which accounts for most of the gain; a small comparison head adds a moderate improvement, and neither component has retriever-specific parameters. RAGCompass improves answer F1 by 28% relative to the best fixed retriever, and outperforms fusion baselines, three query-only routers, and a prompted 8B language model. Generalization experiments show that RAGCompass keeps most of its gain under an unseen corpus, retriever, or reader, while query-only baselines fall to the best fixed retriever or below.

0.11

Finding the Right Balance: Relevance and Diversity in LLM Retrieval

cs.IR cs.CL ·2026-10-07 ·Guillaume Brouillette, Faustin Kagabo, Usef Faghihi, Nadia Ghazzali
code/data similarity 0.63
Abstract · 36 pages, 8 figures, 13 tables. Code and results: https://github.com/GuillaumeBrouillette/finding-the-right-balance

Retrieval diversification is widely available in retrieval-augmented generation (RAG) frameworks, yet prior studies disagree on whether it improves retrieval and answer quality. We show that its effectiveness varies primarily with candidate-pool redundancy, in a pattern consistent with the number of distinct evidence pieces a query requires. Using controlled near-duplicate injection and production-style overlapping chunking, we find that diversification harms relevance, evidence coverage and answer quality on clean pools, but becomes beneficial on multi-evidence tasks when redundancy causes nearest-neighbor retrieval to select repeated passages. We therefore introduce a query-adaptive rule that diversifies only when the effective number of distinct documents in the nearest-neighbor top-$k$ selection falls below the query's evidence requirement. Computed from existing embeddings, the rule captures most of the achievable gain, transfers across datasets and encoders and automatically reduces to nearest-neighbor retrieval for single-evidence queries. We also introduce RNG-Score, a geometric reranker with an exact nearest-neighbor fallback whose margin indicates duplicate structure. Overall, we conclude that diversification should be used selectively, based on observable redundancy and evidence requirements.

0.30

Fidelity Synopses: Pre-Execution Plan Selection for Approximate Vector Search

cs.DB cs.IR ·2026-10-30 ·Adithya Ramanatha, Sanskar S. Singh, Maya Ramanath
benchmark similarity 0.62
Abstract

Vector search has become a database operator with many physical implementations: different index families, build-time and search-time parameters, and compressed or disk-resident variants. Each trades cost against fidelity, the fraction of a query's exact top-k nearest neighbors that a configuration returns. Deployments nevertheless choose one configuration offline and use it for every query, so they are sized for their hardest queries and overpay for the rest. On a 5.8M-item LAION benchmark with four embedding models, running each query on the cheapest configuration that suffices for it would roughly halve the cost in a typical setting. Exploiting this requires predicting each configuration's fidelity before running it, which is hard because fidelity is defined by the very exact search that approximate indexes avoid. Relational optimizers solve the analogous problem for cardinality with offline synopses; we propose one for vector search. A fidelity synopsis divides the embedding space into cells around a set of landmarks and records, for every cell and configuration, how often that configuration met a fidelity target on a sample of queries. At query time, one nearest-landmark lookup yields a calibrated estimate for every configuration, and an optimizer chooses configurations so that the workload misses the target no more often than its requirement allows. We test the two premises the synopsis relies on, that fidelity varies smoothly across the embedding space and that its loss is specific to each configuration. Held to the same requirement a static deployment is built to meet, synopsis-guided selection lowers cost in all 16 combinations of model and query direction, to a median of 0.72x the static deployment, with a batched lookup of 21-32 microseconds per query. The saving persists against the best static choice among the same indexes, and within a single index at no extra memory.

0.09

Preference Is Not Position: Evaluating Reranking on NevIR in Fixed Hybrid Candidate Pools

cs.IR ·2026-10-19 ·Yuning Bian, Lele Xu, Yifan Yuan, Xinyu Yin, Chengyan Zhan
publishedcode/data similarity 0.62
Abstract · 6 pages, 2 figures, 3 tables. Accepted at the IEEE BigData 2026 Undergraduate and REU Consortium. Code: https://github.com/ql-link/LinkRag-E…

Near-identical passages can differ in a decisive condition, such as negation. A ranker may order such a pair correctly while placing the preferred passage far down the retrieved list. We examine this distinction using official NevIR pairs jointly covered by fixed candidate pools from dense, learned sparse, and BM25 retrieval. On 2,736 Test query directions, a pairwise-trained ranker improves strict per-query preference over fixed fusion (56.62% versus 52.30%) while reducing preferred-passage mean reciprocal rank (MRR) from 0.5764 to 0.3055. An exploratory fixed-model replay of a variant with background sampling and revised supervision yields MRR of 0.5859 and strict preference of 54.53%. Bounded pointwise reranking of the fusion top 20 improves both outcomes. Qwen3-14B-AWQ with fusion tie-breaking reaches 77.45% strict preference, 58.99% NevIR-style paired accuracy across 1,363 complete pairs, and preferred-passage MRR of 0.7509. A BGE cross-encoder reaches 69.77% strict preference and MRR of 0.7404. A post-hoc analysis finds that the branch where equal Qwen grades for the two designated passages trigger fusion tie-breaking accounts for 27% of evaluated directions and contains most remaining strict-preference failures. These findings support joint evaluation of pairwise preference, designated-passage position, and computational cost, while distinguishing model judgments from the combined ranking procedure.

0.50