Multi-query Multi-passage Late Fusion (MMLF)

• MMLF is defined as a method that expands a user query into several sub-queries, generates pseudo-passages, and fuses independent ranking lists using techniques like Reciprocal Rank Fusion.
• It integrates both lexical and dense retrieval signals through delayed fusion, where multi-query and multi-passage evidence is aggregated after individual scoring.
• Empirical studies show that MMLF consistently improves retrieval metrics across diverse settings by leveraging multiple query perspectives and scalable late fusion techniques.

Multi-query Multi-passage Late Fusion (MMLF) is a retrieval design in which a system constructs multiple views of a query, evaluates them against multiple passage-like retrieval units or passage-conditioned representations, and delays combination until after independent scoring or ranking. In the explicit formulation used within domain-aware RAG, a user query $q$ is expanded into sub-queries $\mathcal{Q}=\{q_1,\dots,q_n\}$, a pseudo-document $p_i$ is generated for each $q_i$, and the ranking lists produced from the raw query and the pseudo-documents are fused with Reciprocal Rank Fusion (RRF) to yield the final coarse-ranking output (Lin et al., 8 Sep 2025). Related conversational retrieval work places this design in a broader late-fusion family: conversational multi-query rewriting supplies weighted query variants that can feed MMLF-style systems (Kostric et al., 2024), and conversational-memory retrieval demonstrates that late interaction over turn-level passages plus lexical–dense score fusion is substantially stronger than early pooling at session level (Lysenstøen, 2 Jun 2026).

1. Formal definition and retrieval scope

In MoLER, MMLF is defined operationally as a three-step process: instruction expansion, pre-answer guidance, and reciprocal rank fusion. Given a user query $q$, instruction expansion produces $n$ sub-queries,

$\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$

and pre-answer guidance generates a distinct pseudo-passage $p_i$ for each $q_i$. Retrieval is then run independently for the original query and for each pseudo-passage, producing lists $L_0,L_1,\dots,L_n$, which are fused by

$\mathcal{Q}=\{q_1,\dots,q_n\}$0

with $\mathcal{Q}=\{q_1,\dots,q_n\}$1 (Lin et al., 8 Sep 2025).

This definition is broad enough to cover several retrieval settings. In conversational passage retrieval, the underlying information need is conditioned on dialogue history $\mathcal{Q}=\{q_1,\dots,q_n\}$2, and the goal is to retrieve passages $\mathcal{Q}=\{q_1,\dots,q_n\}$3 relevant to the current utterance $\mathcal{Q}=\{q_1,\dots,q_n\}$4 (Kostric et al., 2024). In conversational-memory retrieval, the retrieval unit is a session $\mathcal{Q}=\{q_1,\dots,q_n\}$5, where each turn $\mathcal{Q}=\{q_1,\dots,q_n\}$6 functions as a passage and session-level ranking is therefore a multi-passage problem (Lysenstøen, 2 Jun 2026). A plausible implication is that MMLF is best understood not as a single model family, but as a retrieval pattern: multiple query views, multiple passage-level evidence sources, and delayed aggregation.

2. Multi-query construction

The multi-query component of MMLF is implemented differently across the cited systems, but the common objective is to preserve uncertainty and aspect diversity instead of collapsing everything into a single rewrite. In MoLER, the multi-query stage is called Multi Query Retriever (MQR): for a given query $\mathcal{Q}=\{q_1,\dots,q_n\}$7, the policy $\mathcal{Q}=\{q_1,\dots,q_n\}$8 generates $\mathcal{Q}=\{q_1,\dots,q_n\}$9 expanded sub-queries under an MQR prompt, with $p_i$0 used by default in most experiments (Lin et al., 8 Sep 2025).

Conversational passage retrieval provides a distinct but closely related mechanism. A T5-base sequence-to-sequence query rewriter is run with beam search of width $p_i$1, and instead of retaining only the top-1 beam, the system keeps the top $p_i$2 beams, usually $p_i$3. Each rewrite $p_i$4 receives a length-normalized sequence score

$p_i$5

and these scores are later used as fusion weights (Kostric et al., 2024). The paper emphasizes that, relative to a beam-search baseline already using $p_i$6, retaining all beams incurs “no additional cost” because the beam-search computation has already been performed.

These two constructions instantiate different philosophies. MQR uses prompting to produce diverse task-oriented sub-queries (Lin et al., 8 Sep 2025), whereas CMQR extracts diversity from the beam structure of a trained rewriter and makes the rewrite probabilities explicit (Kostric et al., 2024). This suggests that MMLF is agnostic to the origin of query multiplicity: query views may come from prompting, sequence-generation uncertainty, or other reformulation mechanisms, provided they can be scored or fused coherently.

3. Multi-passage construction and late-fusion operators

The “multi-passage” component likewise admits multiple realizations. In MoLER, each expanded query $p_i$7 is paired with its own pseudo-passage,

$p_i$8

and retrieval is performed with a fixed embedding model, OpenAI text-embedding-ada-002, using cosine similarity between the embedding of $p_i$9 or $q_i$0 and the embedding of each document (Lin et al., 8 Sep 2025). Here the passage objects are not corpus fragments but generated pseudo-documents that act as retrieval probes.

Conversational-memory retrieval studies a different multi-passage regime in which a session is the retrieval unit and individual turns are the passages. With a frozen bi-encoder $q_i$1, turn isolation retrieval scores each session by aggregating per-turn query–turn similarities. The main operators examined are

$q_i$2

and

$q_i$3

with max-sim and top-$q_i$4 both outperforming early interaction based on mean-pooled session embeddings (Lysenstøen, 2 Jun 2026). The same work then performs a second late fusion step at score level, combining BM25 and dense late-interaction scores via query-wise $q_i$5-normalization: $q_i$6

A central technical point is that late fusion is not tied to one aggregation operator. RRF combines ranked lists in MoLER (Lin et al., 8 Sep 2025); max-sim and top-$q_i$7 combine per-passage dense scores in conversational memory retrieval (Lysenstøen, 2 Jun 2026). By contrast, fixed-parameter smooth-max was shown to be brittle: with $q_i$8, log-sum-exp collapsed for half of the tested encoders on LoCoMo (Lysenstøen, 2 Jun 2026). The evidence therefore favors scale-free passage aggregators such as max or top-$q_i$9, or rank-based fusion such as RRF, over uncalibrated smooth pooling.

4. Variants, neighboring formulations, and architectural relations

The phrase MMLF is used most explicitly in MoLER, but several adjacent formulations occupy nearby points in the same design space.

Formulation	Multi-query component	Fusion locus
CMQR	Top-$q$0 query rewrites from beam search	Term weights or query embeddings
Conversational-memory late interaction	One query per question	Turn-to-session aggregation, then lexical–dense score fusion
MoLER MMLF	MQR sub-queries plus one pseudo-passage per sub-query	Rank-level RRF over $q$1

CMQR is conceptually similar to MMLF but not identical. In its sparse form, all rewrites are merged into a single weighted bag-of-words query, with raw term weights

$q$2

followed by normalization across terms. In its dense form, rewrite embeddings are merged into a weighted centroid,

$q$3

and retrieval proceeds with a single query vector (Kostric et al., 2024). The paper explicitly states that there is no rank-based fusion stage; fusion occurs earlier, at the term or embedding level. A common misconception is therefore that “late fusion” in all MMLF-adjacent work means merging ranked lists. The conversational rewriting paper shows that the same multi-query signal can be fused before retrieval rather than after it.

MoLER introduces another distinction through Multi-query Single-passage Late Fusion (MSLF), its training-time surrogate for MMLF. In MSLF, the system still generates multiple sub-queries $q$4, but CQE is called only once with all queries to produce a single pseudo-passage $q$5, and RRF is then applied to two rank lists, from $q$6 and $q$7. This reduces the number of model interactions from $q$8 to $q$9 during policy rollouts. At inference, the same policy $n$0 switches to full MMLF, generating one pseudo-passage per sub-query and fusing all resulting lists (Lin et al., 8 Sep 2025). This training–inference asymmetry is central to MoLER’s efficiency claims.

5. Empirical behavior

The empirical literature around MMLF and closely related late-fusion systems shows three recurring patterns: multi-query evidence improves retrieval, multi-passage late interaction is materially better than early pooling, and lexical–dense complementarity is substantial but domain-dependent.

In conversational passage retrieval on QReCC, CMQR improved sparse BM25 overall MRR from 31.03 to 37.34 for T5QR$n$1, from 49.18 to 50.24 for ConvGQR, and from 46.72 to 50.73 for T5QR$n$2; in dense GTR retrieval, the corresponding MRR improvements were from 36.08 to 40.53, from 42.18 to 45.82, and from 42.46 to 45.98. The reported gains versus the single-query counterparts were statistically significant at $n$3 (Kostric et al., 2024). These are not full MMLF results, but they show that retaining multiple rewrite views reliably improves first-pass retrieval.

In conversational-memory retrieval on LoCoMo, the late-minus-early gap was large for all six tested encoders. For e5-large-v2, dense early interaction achieved Hit@1 $n$4, dense late max-sim achieved $n$5, BM25 achieved $n$6, and BM25 plus late dense fusion achieved Hit@1 $n$7 and NDCG@5 $n$8. Across six encoders, score-level fusion of BM25 with late interaction improved LoCoMo Hit@1 by $n$9 to $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$0 points over late interaction alone, all with $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$1 (Lysenstøen, 2 Jun 2026). The same paper also found that reranking the fused top-10 with ms-marco-MiniLM-L-6-v2 degraded Hit@1 from $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$2 to $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$3, a drop of $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$4 points, indicating that a generic web-search cross-encoder can be harmful in this domain.

The gains are not uniform across datasets. On LongMemEval-S, BM25 already reached R@5 $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$5 and NDCG@5 $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$6; dense late interaction gave R@1 $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$7, and BM25 plus late fusion gave R@1 $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$8, only $\mathcal{Q}=\{q_1,q_2,\dots,q_n\},$9 over BM25, with a confidence interval including zero and $p_i$0 (Lysenstøen, 2 Jun 2026). This suggests that when lexical overlap saturates the task, sophisticated dense multi-passage fusion adds comparatively little.

MoLER reports that MMLF consistently outperforms its cheaper MSLF training-time analogue at inference. For Qwen3-1.7B with MoL and Dr.GRPO on SCIFACT, MSLF obtained Recall@10 $p_i$1 and nDCG@10 $p_i$2, whereas MMLF reached Recall@10 $p_i$3 and nDCG@10 $p_i$4. In the main results table, Qwen3-1.7B+MoL+GRPO achieved NFCORPUS Recall@1k $p_i$5 and SCIFACT Recall@10 $p_i$6, compared with $p_i$7 and $p_i$8 for Qwen3-32B+MMLF (Lin et al., 8 Sep 2025). The paper further reports nearly logarithmic scaling of recall with the number of query expansions in MMLF, whereas MSLF scales poorly and plateaus quickly.

6. Limitations, misconceptions, and open problems

One misconception is that MMLF necessarily requires independent retrieval for every query variant followed by rank fusion. The CMQR results show a different possibility: multi-query evidence can be fused earlier, either as sparse term weights or as a dense centroid embedding, while still preserving the underlying rationale of combining multiple query perspectives (Kostric et al., 2024). Another misconception is that adding a reranker is automatically beneficial. The conversational-memory evidence shows that at least one off-the-shelf cross-encoder reranker can degrade a strong fused first stage by $p_i$9 points of Hit@1 (Lysenstøen, 2 Jun 2026).

The current literature also imposes clear scope limits. CMQR is confined to first-pass retrieval and explicitly identifies multi-query use in re-ranking as future work (Kostric et al., 2024). The conversational-memory study is a retrieval-stage account rather than a complete memory architecture, and it does not test multi-query generation per question, even though its equations are compatible with that extension (Lysenstøen, 2 Jun 2026). MoLER depends on domain corpora for continual pre-training, relevant-document labels for RL reward construction, and a multi-step pipeline with multiple LLM calls, multiple retrievals, and RRF fusion; the paper also notes diminishing returns as $q_i$0 grows and acknowledges the risk of off-topic expansions or hallucinated pseudo-passages (Lin et al., 8 Sep 2025).

The open research directions are correspondingly structured. The cited works propose or motivate adaptive selection of the number of rewrites, multi-query use in re-ranking components, query-type-dependent lexical–dense weights $q_i$1, and better-calibrated smooth pooling operators (Kostric et al., 2024, Lysenstøen, 2 Jun 2026). MoLER points toward training regimes in which an efficient surrogate such as MSLF is used during RL, while full MMLF is reserved for inference, and toward broader integration of domain-aware continual pre-training, GRPO-style optimization, and scalable late-fusion retrieval (Lin et al., 8 Sep 2025). Taken together, these results position MMLF less as a single fixed algorithm than as a family of retrieval architectures organized around three principles: multiplicity of query views, preservation of passage-level evidence, and deferred combination of heterogeneous retrieval signals.