---
title: 'BM25-V: Variants and Interpretations in IR'
url: https://www.emergentmind.com/topics/bm25-v
type: topic
---

# BM25-V: Variants and Interpretations in IR

Searching arXiv for BM25-V and closely related usages to ground the article in recent papers.
arxiv_search(query="BM25-V OR \"BM25-V\" OR \"BM25 V\" OR \"semantic variant of BM25\" OR \"visual words meet BM25\"", max_results=10, sort_by="submittedDate")
BM25-V is a non-standard label used in recent retrieval literature for several technically distinct constructions. Taken together, these works suggest not a single canonical algorithm but a family of BM25-related ideas that preserve some combination of term-frequency-like evidence, inverse document frequency, length normalization, sparse inverted-index retrieval, or BM25-compatible interpretability while altering the representation being scored, the model being explained, or even the retrieval modality itself [2304.12631] [2502.04645] [2301.09728] [2305.14087] [2603.05781]. In some contexts, BM25-V denotes a sparse surrogate explanation of dense ranking; in others, a semantic BM25-like circuit identified inside a cross-encoder; in others still, a learned query-side augmentation of BM25 or a visual-word image retriever. Conversely, some recent BM25-based systems explicitly introduce no BM25-V variant at all [2410.04620].

## 1. Terminological status and major senses

The literature does not treat BM25-V as a standardized IR term with a single agreed formula. Instead, the same label has been attached to multiple different methods, each retaining a recognizable BM25 core while changing what is matched, how the query is represented, or what explanatory role BM25 plays.

| Sense of BM25-V | Core mechanism | Representative paper |
|---|---|---|
| Sparse surrogate explanation | Optimize an equivalent sparse query so BM25 approximates a dense ranked list | [2304.12631] |
| Semantic BM25 circuit | Identify soft-TF-, saturation-, length-, and IDF-like components inside a cross-encoder | [2502.04645] |
| BM25 score injection | Insert the BM25 score as text into a cross-encoder input | [2301.09728] |
| Learned sparse query augmentation | Predict sparse query expansion and re-weighting for BM25 end-to-end | [2305.14087] |
| Visual-word retrieval | Apply Okapi BM25 to SAE-derived visual words from ViT patch features | [2603.05781] |

This multiplicity matters because the label can easily be misread as referring to a single retrieval formula. Recent work on Polish passage retrieval, for example, uses standard OKAPI BM25 as a first-stage retriever and explicitly does not introduce any BM25-V or custom BM25 formulation; BM25 is used conventionally before reranking with an ensemble of cross-encoders [2410.04620]. A careful reading of context is therefore necessary whenever the term appears.

## 2. BM25-V as a sparse surrogate for dense ranking

In "Explain like I am BM25," BM25-V is best understood as a local interpretability method for dense retrieval rather than a new first-stage retriever [2304.12631]. The central object is an equivalent query \(Q^+\): a sparse query over the vocabulary extracted from the dense model’s top-\(k\) documents, chosen so that BM25 reproduces the dense model’s ranked list as closely as possible. The optimization target is

$$
\arg\max_{Q^+ \subset V(L_k(Q; \theta))} \omega(L_k(Q^+; \phi), L_k(Q; \theta)),
$$

where \(L_k(Q;\theta)\) is the dense model’s top-\(k\) list, \(\phi\) is the sparse retriever, and \(\omega\) is a ranked-list similarity measure. The paper uses Jaccard and especially Rank-Biased Overlap.

This formulation differs from pseudo-relevance feedback and RM3 in two ways. First, the objective is explicitly ranked-list fidelity to a target neural model, not generic expansion quality. Second, the resulting equivalent query may omit original query terms, because fidelity to the dense model’s behavior takes precedence over preserving the original wording. BM25 becomes an interpretable sparse surrogate whose vocabulary-level behavior serves as a verbalization of the dense model’s implicit concept set.

Because the search over candidate sparse queries is NP-complete, the paper uses a best-first-search-style discrete exploration rather than exact optimization. The search starts from the empty query, explores states obtained by adding or removing terms, restricts the candidate vocabulary to the dense model’s top-\(k\) documents, favors additions according to normalized RM3 weights, and discourages removing high-tf-idf terms. The method is evaluated on MS MARCO passage ranking with TREC DL 2019 topics and targets including ANCE, ColBERT, MonoT5, DeepCT + ColBERT, and ColBERT + BERT-based query expansion [2304.12631].

The reported fidelity and effectiveness are substantial. BFS-based equivalent queries achieve RBO up to \(0.5194\) and Jaccard up to \(0.5327\) for MonoT5, and BM25 run with the generated \(Q^+\) can retain up to about \(96\%\) of the target model’s nDCG. The same paper also reports an average latency of about \(6\) seconds and notes that failure cases arise when the vocabulary of the dense model’s top-\(k\) neighborhood is too limited. BM25-V in this sense is therefore a local, approximate, query-level explanation method: human-readable, model-specific, and tied to ranked-list agreement rather than internal token attributions.

## 3. BM25-V as a semantic BM25 circuit inside a cross-encoder

A very different use of the term appears in "Cross-Encoder Rediscovers a Semantic Variant of BM25" [2502.04645]. Here BM25-V is not introduced as a separate benchmark algorithm; rather, it names the mechanistic claim that a MiniLM cross-encoder implements a semantic variant of BM25. The paper argues that the model computes the classical BM25 ingredients—term frequency, term saturation, document length normalization, and inverse document frequency—but in soft semantic form rather than through exact lexical counts.

The proposed circuit has localized functional components. Early-layer Matching Heads compute a Matching Score behaving like semantic TF while also reflecting term saturation and document-length effects. Two middle-layer Contextual Query Representation Heads, 8.10 and 9.11, redistribute soft-TF information from higher-IDF query terms across the query representation. Four layer-10 Relevance Scoring Heads, 10.1, 10.4, 10.7, and 10.10, read out those signals in an IDF-sensitive way. The paper further identifies a dominant rank-1 embedding component, \(U_0\), whose values correlate about \(-71.36\%\) with MS MARCO IDF values, leading the authors to interpret it as a one-dimensional IDF dictionary [2502.04645].

The empirical support comes from path patching, attention-pattern analysis, axiom-based diagnostics, ablation, and linear reconstruction. On the TFC1 and STMC1 diagnostic datasets, the patching importance values for the identified relevance heads have correlation \(= 0.9996\), \(p < 0.001\), which the paper interprets as evidence that exact and semantic matches are processed similarly. Matching Heads show much stronger correlation between attention and semantic similarity than other heads, with average Pearson correlation \(0.500\) versus \(0.132\). Mean-ablation of Matching Heads causes large score drops: on TFC-perturbed samples, the average value drops from \(5.146\) to \(-4.394\); on STMC-perturbed samples, from \(-1.998\) to \(-4.611\). A linear regression using \(U_0(q_i)\) and Matching Scores reaches Pearson correlation \(0.8157\) with the cross-encoder’s relevance scores, exceeding BM25 under the tuned parameters at \(0.4200\); across 12 IR datasets, the paper reports median Pearson \(0.8401\), median Spearman \(0.7619\), and NDCG@10 alignment \(88.4\%\) [2502.04645].

The significance of this usage is interpretive rather than procedural. BM25-V denotes a mechanistically localized semantic scoring circuit, which suggests that at least one cross-encoder has not simply learned an arbitrary ranking function but has rediscovered BM25-like relevance computation in a semantically softened form.

## 4. Query-side BM25-V constructions

Two additional lines of work use BM25-V to denote explicitly query-side modifications rather than document-side or model-internal reinterpretations.

In "Injecting the BM25 Score as Text Improves BERT-Based Re-rankers," the relevant model family is CE\(_{\text{BM25CAT}}\), which inserts the BM25 score directly into the cross-encoder input as text [2301.09728]. The standard cross-encoder input
\([ \mathrm{CLS}]\, q \,[\mathrm{SEP}]\, p \,[\mathrm{SEP}]\)
is changed to
\([ \mathrm{CLS}]\, q \,[\mathrm{SEP}]\, \text{BM25} \,[\mathrm{SEP}]\, p \,[\mathrm{SEP}]\).
The paper studies several representations of the score and finds that global Min-Max normalization followed by integer conversion works best. On MSMARCO dev, the injected model improves all four tested cross-encoders: BERT-Base rises from MRR@10 \(=.342\) to \(=.364\), BERT-Large from \(=.360\) to \(=.369\), DistilBERT from \(=.325\) to \(=.339\), and MiniLM from \(=.360\) to \(=.367\). In the exact-matching probe where non-query passage words are masked, BM25 scores \(=.187\), BERT-Base CAT scores \(=.180\), and BERT-Base BM25CAT scores \(=.202\). Integrated Gradients also assigns the BM25 token high attribution, with mode rank \(3\), indicating that the model actively uses the injected lexical score [2301.09728].

In "BM25 Query Augmentation Learned End-to-End," BM25-V refers to a learned sparse query expansion and re-weighting framework that preserves BM25’s inverted-index behavior [2305.14087]. Standard BM25 is written as \(\bigl(v \odot \mathrm{bow}(q)\bigr)^\top f(d)\), where \(v\) is the IDF vector and \(f(d)\) is the BM25 document term-frequency vector. The learned model predicts an augmentation vector \(a(q)\) and a term-weight vector \(w(q)\), producing the score

$$
\mathrm{score}(q,d) = \Bigl( w(q) \odot v \odot \bigl(\mathrm{bow}(q)+a(q)\bigr) \Bigr)^\top f(d).
$$

The augmentation is learned end-to-end with a contrastive retrieval objective and a frequency-weighted sparsity regularizer. Using `distilbert-base-uncased`, the method improves over BM25 while staying close to BM25 latency: on Natural Questions, Acc@5 rises from \(0.436\) for BM25 (Pyserini) to \(0.557\); on EntityQuestions, from \(0.616\) to \(0.693\); on MSMARCO, NDCG@10 rises from \(0.228\) to \(0.251\) and Recall@100 from \(0.658\) to \(0.687\), with latency \(0.030\)s versus \(0.020\)s for BM25 (Pyserini) [2305.14087].

These two query-side constructions are methodologically unrelated, but both retain BM25 as a sparse lexical substrate while giving the query additional structure. One exposes BM25 to the reranker as an explicit tokenized numeric signal; the other learns a richer sparse query representation before retrieval.

## 5. BM25-V beyond lexical text: visual and latent vocabularies

The most explicit non-textual instantiation appears in "Visual Words Meet BM25: Sparse Auto-Encoder Visual Word Scoring for Image Retrieval" [2603.05781]. BM25-V here is a sparse image retrieval method that applies Okapi BM25 to visual words derived from SAE activations on frozen ViT patch features. The backbone uses final-layer SigLIP2 patch features \(\mathbf{Z}^{(26)} \in \mathbb{R}^{P \times D}\) with \(P = 729\) and \(D = 1152\). An SAE with expansion factor \(e=16\) yields a visual vocabulary of \(18{,}432\) dimensions; patch-level sparsity is \(k=16\), and post-pooling sparsity is also \(k_{\text{post}}=16\). The pooled sparse image vector is treated as a bag of visual words, with term frequency given by visual-word activation, document length by \(\|\mathbf{v}\|_1\), and document frequency by the number of reference images activating the corresponding SAE dimension [2603.05781].

The key empirical justification is that these visual words exhibit a Zipfian-like document-frequency distribution. Across seven datasets, the paper reports power-law fits with \(R^2 \in [0.917, 0.973]\) and exponents \(\alpha \in [1.20, 2.32]\). It also notes that head dimensions fire in more than \(50\%\) of images, only about \(0.25\%\)–\(1.83\%\) of active dimensions are stop-word-like pervasive dimensions, and \(96\%\)–\(99.7\%\) of active dimensions have IDF \(> 2\). BM25-V serves as the first stage of a two-stage system: it retrieves candidates sparsely, then reranks the top \(K=200\) with dense cosine similarity from the same frozen backbone. First-stage recall is high—Recall@100 ranges from \(0.984\) to \(0.999\), and Recall@200 is at least \(0.993\) across all seven benchmarks. After reranking, average R@1 is \(0.857\) versus a dense baseline of \(0.859\), a drop of about \(0.2\) percentage points on average; on DTD and Flowers-102, the two-stage system improves dense retrieval by \(+0.7\%\) and \(+0.1\%\), respectively. The sparse representation is also reported as about \(48\times\) smaller than float32 dense embeddings when \(L_0=16\) [2603.05781].

A closely related but separately named development is "Latent Terms," which shows that dense retrievers contain trivially extractable BM25-ready latent vocabularies [2605.29384]. There, a frozen dense retriever is paired with a Sparse Autoencoder of latent vocabulary size \(32{,}768\) and top-\(k\) sparsity \(16\), trained only with reconstruction loss on unlabeled text. The resulting latent features have approximately Zipfian collection statistics and can be scored effectively with BM25 after sum-pooling and a square-root transform. A plausible implication is that the visual-word formulation of BM25-V belongs to a broader movement in which BM25 is increasingly applied to learned sparse vocabularies rather than only to surface words.

## 6. Relation to standard BM25, neighboring variants, and recurrent misconceptions

One recurrent misconception is that BM25-V denotes any modern BM25-based retrieval pipeline. The literature does not support that reading. In the Poleval 2023 winning system for Polish passage retrieval, BM25 is simply standard OKAPI BM25 with \(k_1 = 1.2\), \(b = 0.75\), and \(\varepsilon = 0.25\), preceded by tokenization with `nltk.tokenize.word_tokenize`, lowercasing, stemming with `pystempel` using the Polimorf stemmer, and Polish stopword removal [2410.04620]. BM25 is used both for inference-time candidate generation and for drawing \(100\) negative query-passage pairs per positive passage from the top \(2000\) BM25 results during reranker training. Candidate pools are domain-specific—top \(3000\) for wiki-trivia, top \(1500\) for legal-questions, and all \(921\) passages for allegro-faq—and BM25 alone scores \(52.67\) NDCG@10 on test-A and \(42.55\) on test-B, while the final cross-encoder ensemble reaches \(75.40\) and \(69.36\), respectively. The paper explicitly does not propose a BM25-V variant [2410.04620].

Another source of ambiguity is that other genuine BM25 modifications are usually named differently. On BRIGHT, "Lighting the Way for BRIGHT" shows that a major reproduction discrepancy comes from query representation: standard Anserini/Pyserini BM25 uses a bag-of-words query vector, whereas BRIGHT’s published baseline applies BM25 weighting on the query side itself [2509.02558]. The overall average nDCG@10 is \(0.137\) for BoW, \(0.135\) for BoW accurate, \(0.148\) for query-side BM25, \(0.145\) for query-side BM25 accurate, and \(0.145\) for BRIGHT original BM25. The paper concludes that query-side weighting matters more than Lucene’s approximate versus exact length normalization in this long-query regime, especially for roughly \(16\)–\(256\)-token queries. This is a meaningful BM25 variant, but it is called query-side BM25, not BM25-V [2509.02558].

Likewise, "BMX" is an explicit extension of BM25 rather than a BM25-V paper [2408.06643]. It augments the BM25 TF-IDF backbone with entropy-weighted similarity and semantic enhancement through weighted query augmentation. The method introduces parameters \(\alpha\) and \(\beta\), a token entropy term \(\mathrm{E}(q_i)\), a similarity term \(\mathrm{S}(Q,D)\), and an augmented-query scoring function that aggregates the original query with LLM-generated related queries. BM\(\mathcal{X}\) is thus a formally named BM25 extension in its own right, not a synonym for BM25-V [2408.06643].

The overall lesson is terminological as much as technical. BM25-V is best treated as a context-dependent label whose meaning must be recovered from the specific paper: sparse surrogate explanation, semantic BM25 circuit, query-side neural augmentation, or visual-word retrieval are all attested uses, whereas many strong BM25 systems and several important BM25 variants do not use the term at all.

Source: https://www.emergentmind.com/topics/bm25-v