BM25-V: Variants and Interpretations in IR

• BM25-V is a term describing a family of BM25-derived retrieval methods that vary by altering representations, scoring mechanisms, or modalities.
• These methods retain core features such as term frequency, inverse document frequency, and length normalization while incorporating semantic interpretation and query augmentation.
• BM25-V offers flexible, interpretable retrieval approaches with high fidelity to dense models, applicable across text, image, and latent feature retrieval.

Searching arXiv for BM^^^^4 OR \4^^^^5-V and closely related usages to ground the article in papers. arxiv_search(query="^^^^^^^^4BM25-V OR \4^^^^^^^^"BM^^^^4 OR \4^^^^5-V\"^^^^^^^^4 OR \4^^^^^^^^"BM^^^^4 OR \4^^^^5 V\"^^^^^^^^4 OR \4^^^^^^^^"semantic variant of BM^^^^4 OR \4^^^^5\"^^^^^^^^4 OR \4^^^^^^^^"visual words meet BM^^^^4 OR \4^^^^5\"", max_results=^^^^4 OR \4BM25-V OR \4^^^^, sort_by="submittedDate") BM^^^^4 OR \4^^^^5-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 BM^^^^4 OR \4^^^^5-related ideas that preserve some combination of term-frequency-like evidence, inverse document frequency, length normalization, sparse inverted-index retrieval, or BM^^^^4 OR \4^^^^5-compatible interpretability while altering the representation being scored, the model being explained, or even the retrieval modality itself (&&&^^^^4BM25-V OR \4^^^^&&&, &&&^^^^4 OR \4^^^^&&&, &&&^^^^4 OR \4^^^^&&&, &&&^^^^4 OR \4^^^^&&&, Han et al., 6 Mar 2026). In some contexts, BM^^^^4 OR \4^^^^5-V denotes a sparse surrogate explanation of dense ranking; in others, a semantic BM^^^^4 OR \4^^^^5-like circuit identified inside a cross-encoder; in others still, a learned query-side augmentation of BM^^^^4 OR \4^^^^5 or a visual-word image retriever. Conversely, some recent BM^^^^4 OR \4^^^^5-based systems explicitly introduce no BM^^^^4 OR \4^^^^5-V variant at all (Pokrywka, 2024).

^^^^4 OR \4^^^^. Terminological status and major senses

The literature does not treat BM^^^^4 OR \4^^^^5-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 BM^^^^4 OR \4^^^^5 core while changing what is matched, how the query is represented, or what explanatory role BM^^^^4 OR \4^^^^5 plays.

Sense of BM^^^^4 OR \4^^^^5-V	Core mechanism	Representative paper
Sparse surrogate explanation	Optimize an equivalent sparse query so BM^^^^4 OR \4^^^^5 approximates a dense ranked list	(&&&^^^^4BM25-V OR \4^^^^&&&)
Semantic BM^^^^4 OR \4^^^^5 circuit	Identify soft-TF-, saturation-, length-, and IDF-like components inside a cross-encoder	(&&&^^^^4 OR \4^^^^&&&)
BM^^^^4 OR \4^^^^5 score injection	Insert the BM^^^^4 OR \4^^^^5 score as text into a cross-encoder input	(&&&^^^^4 OR \4^^^^&&&)
Learned sparse query augmentation	Predict sparse query expansion and re-weighting for BM^^^^4 OR \4^^^^5 end-to-end	(&&&^^^^4 OR \4^^^^&&&)
Visual-word retrieval	Apply Okapi BM^^^^4 OR \4^^^^5 to SAE-derived visual words from ViT patch features	(Han et al., 6 Mar 2026)

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 BM^^^^4 OR \4^^^^5 as a first-stage retriever and explicitly does not introduce any BM^^^^4 OR \4^^^^5-V or custom BM^^^^4 OR \4^^^^5 formulation; BM^^^^4 OR \4^^^^5 is used conventionally before reranking with an ensemble of cross-encoders (Pokrywka, 2024). A careful reading of context is therefore necessary whenever the term appears.

^^^^4 OR \4^^^^. BM^^^^4 OR \4^^^^5-V as a sparse surrogate for dense ranking

In "Explain like I am BM^^^^4 OR \4^^^^5," BM^^^^4 OR \4^^^^5-V is best understood as a local interpretability method for dense retrieval rather than a new first-stage retriever (&&&^^^^4BM25-V OR \4^^^^&&&). The central object is an equivalent query PRESERVED_PLACEHOLDER_^^^^4BM25-V OR \4^^^^: a sparse query over the vocabulary extracted from the dense model’s top-PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^ documents, chosen so that BM^^^^4 OR \4^^^^5 reproduces the dense model’s ranked list as closely as possible. The optimization target is

PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^

where PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^ 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 RM^^^^4 OR \4^^^^ 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. BM^^^^4 OR \4^^^^5 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 RM^^^^4 OR \4^^^^ weights, and discourages removing high-tf-idf terms. The method is evaluated on MS MARCO passage ranking with TREC DL ^^^^4 OR \4BM25-V OR \4 OR \4^^^^9 topics and targets including ANCE, ColBERT, MonoT5, DeepCT + ColBERT, and ColBERT + BERT-based query expansion (&&&^^^^4BM25-V OR \4^^^^&&&).

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 BM^^^^4 OR \4^^^^5 run with the generated PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4^^^^ can retain up to about PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^ of the target model’s nDCG. The same paper also reports an average latency of about PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^ seconds and notes that failure cases arise when the vocabulary of the dense model’s top-PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^ neighborhood is too limited. BM^^^^4 OR \4^^^^5-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.

^^^^4 OR \4^^^^. BM^^^^4 OR \4^^^^5-V as a semantic BM^^^^4 OR \4^^^^5 circuit inside a cross-encoder

A very different use of the term appears in "Cross-Encoder Rediscovers a Semantic Variant of BM^^^^4 OR \4^^^^5" (&&&^^^^4 OR \4^^^^&&&). Here BM^^^^4 OR \4^^^^5-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 BM^^^^4 OR \4^^^^5. The paper argues that the model computes the classical BM^^^^4 OR \4^^^^5 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.^^^^4 OR \4BM25-V OR \4^^^^ and 9.^^^^4 OR \4 OR \4^^^^, redistribute soft-TF information from higher-IDF query terms across the query representation. Four layer-^^^^4 OR \4BM25-V OR \4^^^^ Relevance Scoring Heads, ^^^^4 OR \4BM25-V OR \4^^^^.^^^^4 OR \4^^^^, ^^^^4 OR \4BM25-V OR \4^^^^.4, ^^^^4 OR \4BM25-V OR \4^^^^.7, and ^^^^4 OR \4BM25-V OR \4^^^^.^^^^4 OR \4BM25-V OR \4^^^^, read out those signals in an IDF-sensitive way. The paper further identifies a dominant rank-^^^^4 OR \4^^^^ embedding component, PRESERVED_PLACEHOLDER_^^^^4 OR \44^^^^, whose values correlate about PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^5 with MS MARCO IDF values, leading the authors to interpret it as a one-dimensional IDF dictionary (&&&^^^^4 OR \4^^^^&&&).

The empirical support comes from path patching, attention-pattern analysis, axiom-based diagnostics, ablation, and linear reconstruction. On the TFC^^^^4 OR \4^^^^ and STMC^^^^4 OR \4^^^^ diagnostic datasets, the patching importance values for the identified relevance heads have correlation PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^6, PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^7, 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 PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^8 versus PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^9. Mean-ablation of Matching Heads causes large score drops: on TFC-perturbed samples, the average value drops from PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4^^^^ to PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^; on STMC-perturbed samples, from PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^ to PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^. A linear regression using PRESERVED_PLACEHOLDER_^^^^4 OR \44^^^^ and Matching Scores reaches Pearson correlation PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^5 with the cross-encoder’s relevance scores, exceeding BM^^^^4 OR \4^^^^5 under the tuned parameters at PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^6; across ^^^^4 OR \4 OR \4^^^^ IR datasets, the paper reports median Pearson PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^7, median Spearman PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^8, and NDCG@^^^^4 OR \4BM25-V OR \4^^^^ alignment PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^9 (&&&^^^^4 OR \4^^^^&&&).

The significance of this usage is interpretive rather than procedural. BM^^^^4 OR \4^^^^5-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 BM^^^^4 OR \4^^^^5-like relevance computation in a semantically softened form.

4. Query-side BM^^^^4 OR \4^^^^5-V constructions

Two additional lines of work use BM^^^^4 OR \4^^^^5-V to denote explicitly query-side modifications rather than document-side or model-internal reinterpretations.

In "Injecting the BM^^^^4 OR \4^^^^5 Score as Text Improves BERT-Based Re-rankers," the relevant model family is CEPRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4^^^^, which inserts the BM^^^^4 OR \4^^^^5 score directly into the cross-encoder input as text (&&&^^^^4 OR \4^^^^&&&). The standard cross-encoder input

PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^

is changed to PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^. 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@^^^^4 OR \4BM25-V OR \4^^^^ PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4^^^^ to PRESERVED_PLACEHOLDER_^^^^4 OR \44^^^^, BERT-Large from PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^5 to PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^6, DistilBERT from PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^7 to PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^8, and MiniLM from PRESERVED_PLACEHOLDER_^^^^4 OR \4^^^^9 to $k$^^^^4BM25-V OR \4^^^^. In the exact-matching probe where non-query passage words are masked, BM^^^^4 OR \4^^^^5 scores $k$^^^^4 OR \4^^^^, BERT-Base CAT scores $k$^^^^4 OR \4^^^^, and BERT-Base BM^^^^4 OR \4^^^^5CAT scores $k$^^^^4 OR \4^^^^. Integrated Gradients also assigns the BM^^^^4 OR \4^^^^5 token high attribution, with mode rank $k$4, indicating that the model actively uses the injected lexical score (&&&^^^^4 OR \4^^^^&&&).

In "BM^^^^4 OR \4^^^^5 Query Augmentation Learned End-to-End," BM^^^^4 OR \4^^^^5-V refers to a learned sparse query expansion and re-weighting framework that preserves BM^^^^4 OR \4^^^^5’s inverted-index behavior (&&&^^^^4 OR \4^^^^&&&). Standard BM^^^^4 OR \4^^^^5 is written as $k$5, where $k$6 is the IDF vector and $k$7 is the BM^^^^4 OR \4^^^^5 document term-frequency vector. The learned model predicts an augmentation vector $k$8 and a term-weight vector $k$9, producing the score

$\phi$^^^^4BM25-V OR \4^^^^

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 BM^^^^4 OR \4^^^^5 while staying close to BM^^^^4 OR \4^^^^5 latency: on Natural Questions, Acc@5 rises from $\phi$^^^^4 OR \4^^^^ for BM^^^^4 OR \4^^^^5 (Pyserini) to $\phi$^^^^4 OR \4^^^^; on EntityQuestions, from $\phi$^^^^4 OR \4^^^^ to $\phi$4; on MSMARCO, NDCG@^^^^4 OR \4BM25-V OR \4^^^^ rises from $\phi$5 to $\phi$6 and Recall@^^^^4 OR \4BM25-V OR \4BM25-V OR \4^^^^ from $\phi$7 to $\phi$8, with latency $\phi$9s versus $\omega$^^^^4BM25-V OR \4^^^^s for BM^^^^4 OR \4^^^^5 (Pyserini) (&&&^^^^4 OR \4^^^^&&&).

These two query-side constructions are methodologically unrelated, but both retain BM^^^^4 OR \4^^^^5 as a sparse lexical substrate while giving the query additional structure. One exposes BM^^^^4 OR \4^^^^5 to the reranker as an explicit tokenized numeric signal; the other learns a richer sparse query representation before retrieval.

5. BM^^^^4 OR \4^^^^5-V beyond lexical text: visual and latent vocabularies

The most explicit non-textual instantiation appears in "Visual Words Meet BM^^^^4 OR \4^^^^5: Sparse Auto-Encoder Visual Word Scoring for Image Retrieval" (Han et al., 6 Mar 2026). BM^^^^4 OR \4^^^^5-V here is a sparse image retrieval method that applies Okapi BM^^^^4 OR \4^^^^5 to visual words derived from SAE activations on frozen ViT patch features. The backbone uses final-layer SigLIP^^^^4 OR \4^^^^ patch features $\omega$^^^^4 OR \4^^^^ with $\omega$^^^^4 OR \4^^^^ and $\omega$^^^^4 OR \4^^^^. An SAE with expansion factor $\omega$4 yields a visual vocabulary of $\omega$5 dimensions; patch-level sparsity is $\omega$6, and post-pooling sparsity is also $\omega$7. The pooled sparse image vector is treated as a bag of visual words, with term frequency given by visual-word activation, document length by $\omega$8, and document frequency by the number of reference images activating the corresponding SAE dimension (Han et al., 6 Mar 2026).

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 $\omega$9 and exponents $k$^^^^4BM25-V OR \4^^^^. It also notes that head dimensions fire in more than $k$^^^^4 OR \4^^^^ of images, only about $k$^^^^4 OR \4^^^^–kk^^^^4 OR \4^^^^ of active dimensions are stop-word-like pervasive dimensions, and $k$4–$k$5 of active dimensions have IDF $k$6. BM^^^^4 OR \4^^^^5-V serves as the first stage of a two-stage system: it retrieves candidates sparsely, then reranks the top $k$7 with dense cosine similarity from the same frozen backbone. First-stage recall is high—Recall@^^^^4 OR \4BM25-V OR \4BM25-V OR \4^^^^ ranges from $k$8 to $k$9, and Recall@^^^^4 OR \4BM25-V OR \4BM25-V OR \4^^^^ is at least $0.5194$^^^^4BM25-V OR \4^^^^ across all seven benchmarks. After reranking, average R@^^^^4 OR \4^^^^ is $0.5194$^^^^4 OR \4^^^^ versus a dense baseline of $0.5194$^^^^4 OR \4^^^^, a drop of about $0.5194$^^^^4 OR \4^^^^ percentage points on average; on DTD and Flowers-^^^^4 OR \4BM25-V OR \4 OR \4^^^^, the two-stage system improves dense retrieval by $0.5194$4 and $0.5194$5, respectively. The sparse representation is also reported as about $0.5194$6 smaller than float^^^^4 OR \4 OR \4^^^^ dense embeddings when $0.5194$7 (Han et al., 6 Mar 2026).

A closely related but separately named development is "Latent Terms," which shows that dense retrievers contain trivially extractable BM^^^^4 OR \4^^^^5-ready latent vocabularies (&&&^^^^4 OR \44^^^^&&&). There, a frozen dense retriever is paired with a Sparse Autoencoder of latent vocabulary size $0.5194$8 and top-$0.5194$9 sparsity $0.5327$^^^^4BM25-V OR \4^^^^, trained only with reconstruction loss on unlabeled text. The resulting latent features have approximately Zipfian collection statistics and can be scored effectively with BM^^^^4 OR \4^^^^5 after sum-pooling and a square-root transform. A plausible implication is that the visual-word formulation of BM^^^^4 OR \4^^^^5-V belongs to a broader movement in which BM^^^^4 OR \4^^^^5 is increasingly applied to learned sparse vocabularies rather than only to surface words.

6. Relation to standard BM^^^^4 OR \4^^^^5, neighboring variants, and recurrent misconceptions

One recurrent misconception is that BM^^^^4 OR \4^^^^5-V denotes any modern BM^^^^4 OR \4^^^^5-based retrieval pipeline. The literature does not support that reading. In the Poleval ^^^^4 OR \4BM25-V OR \4 OR \4 OR \4^^^^ winning system for Polish passage retrieval, BM^^^^4 OR \4^^^^5 is simply standard OKAPI BM^^^^4 OR \4^^^^5 with $0.5327$^^^^4 OR \4^^^^, $0.5327$^^^^4 OR \4^^^^, and $0.5327$^^^^4 OR \4^^^^, preceded by tokenization with nltk.tokenize.word_tokenize, lowercasing, stemming with pystempel using the Polimorf stemmer, and Polish stopword removal (Pokrywka, 2024). BM^^^^4 OR \4^^^^5 is used both for inference-time candidate generation and for drawing $0.5327$4 negative query-passage pairs per positive passage from the top $0.5327$5 BM^^^^4 OR \4^^^^5 results during reranker training. Candidate pools are domain-specific—top $0.5327$6 for wiki-trivia, top $0.5327$7 for legal-questions, and all $0.5327$8 passages for allegro-faq—and BM^^^^4 OR \4^^^^5 alone scores $0.5327$9 NDCG@^^^^4 OR \4BM25-V OR \4^^^^ on test-A and PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4BM25-V OR \4^^^^ on test-B, while the final cross-encoder ensemble reaches PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4 OR \4^^^^ and PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4 OR \4^^^^, respectively. The paper explicitly does not propose a BM^^^^4 OR \4^^^^5-V variant (Pokrywka, 2024).

Another source of ambiguity is that other genuine BM^^^^4 OR \4^^^^5 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 BM^^^^4 OR \4^^^^5 uses a bag-of-words query vector, whereas BRIGHT’s published baseline applies BM^^^^4 OR \4^^^^5 weighting on the query side itself (&&&^^^^4 OR \4^^^^7&&&). The overall average nDCG@^^^^4 OR \4BM25-V OR \4^^^^ is PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4 OR \4^^^^ for BoW, PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \44^^^^ for BoW accurate, PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4^^^^5 for query-side BM^^^^4 OR \4^^^^5, PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4^^^^6 for query-side BM^^^^4 OR \4^^^^5 accurate, and PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4^^^^7 for BRIGHT original BM^^^^4 OR \4^^^^5. 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 PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4^^^^8–PRESERVED_PLACEHOLDER_^^^^4 OR \4BM25-V OR \4^^^^9-token queries. This is a meaningful BM^^^^4 OR \4^^^^5 variant, but it is called query-side BM^^^^4 OR \4^^^^5, not BM^^^^4 OR \4^^^^5-V (&&&^^^^4 OR \4^^^^7&&&).

Likewise, "BMX" is an explicit extension of BM^^^^4 OR \4^^^^5 rather than a BM^^^^4 OR \4^^^^5-V paper (&&&^^^^4 OR \4^^^^9&&&). It augments the BM^^^^4 OR \4^^^^5 TF-IDF backbone with entropy-weighted similarity and semantic enhancement through weighted query augmentation. The method introduces parameters PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4BM25-V OR \4^^^^ and PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4 OR \4^^^^, a token entropy term PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4 OR \4^^^^, a similarity term PRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \4 OR \4^^^^, and an augmented-query scoring function that aggregates the original query with LLM-generated related queries. BMPRESERVED_PLACEHOLDER_^^^^4 OR \4 OR \44^^^^ is thus a formally named BM^^^^4 OR \4^^^^5 extension in its own right, not a synonym for BM^^^^4 OR \4^^^^5-V (&&&^^^^4 OR \4^^^^9&&&).

The overall lesson is terminological as much as technical. BM^^^^4 OR \4^^^^5-V is best treated as a context-dependent label whose meaning must be recovered from the specific paper: sparse surrogate explanation, semantic BM^^^^4 OR \4^^^^5 circuit, query-side neural augmentation, or visual-word retrieval are all attested uses, whereas many strong BM^^^^4 OR \4^^^^5 systems and several important BM^^^^4 OR \4^^^^5 variants do not use the term at all.