Dual-Path Scenario Retrieval

Updated 6 July 2026

Dual-Path Scenario Retrieval is a retrieval architecture that constructs and fuses two distinct evidence paths to cover complementary aspects of query answering.
It combines diverse methods—such as multilingual translation graphs, dense and graph-based retrieval, or coarse and fine filtering—to mitigate single-path limitations.
Empirical evaluations show improved performance in domains like multilingual QA, medical question answering, and autonomous-driving scenario mining.

Searching arXiv for the cited dual-path retrieval papers to ground the article in current preprints. Dual-path scenario retrieval denotes a class of retrieval architectures in which two distinct evidence paths are constructed for the same query, scenario description, or reasoning state, and are then merged, reranked, or selectively activated before final prediction. In recent arXiv usage, the paired paths have included source-language and English-translation decomposition in multilingual multi-hop question answering, graph-based and dense retrieval in medical question answering, category-based and embedding-based search in braking-scenario mining, query-based and pseudo-context-based retrieval in open-domain RAG, document-level and scenario-level indexing in dense retrieval, familiarity and recollection retrieval in personalized memory systems, and coarse vision-language filtering followed by fine text-trajectory matching in robotic autonomy (Wang et al., 19 Mar 2026, Wan et al., 24 Jun 2026, Wu et al., 17 Jul 2025, Chen et al., 7 Jan 2026, Lee et al., 29 Mar 2025, Zhang et al., 10 Mar 2026, Chen et al., 17 Jan 2026). This suggests a general retrieval pattern rather than a single standardized algorithm.

1. Conceptual scope and terminology

Within contemporary retrieval research, dual-path scenario retrieval generally refers to the use of two complementary retrieval channels whose inductive biases differ. One path usually emphasizes coverage, structure, or robustness, while the other emphasizes semantic fidelity, precision, or adaptability. The combination is intended to mitigate the failure mode of any single retrieval pathway when queries are sparse, ambiguous, multilingual, out-of-distribution, or intrinsically multi-hop (Wan et al., 24 Jun 2026, Chen et al., 7 Jan 2026).

The term should be distinguished from earlier, non-retrieval uses of “Dual-Path.” Saboori et al. use “Dual-Path” to denote an anonymous peer-to-peer communication method with two independent onion-routing chains, a request path and a response path, each with $k=3$ intermediate peers, for a total hop count of $6$ per message (Saboori et al., 2012). That formulation concerns requester anonymity and network reliability rather than evidence retrieval. A related but also distinct usage appears in composed image retrieval, where HINT employs separate reference-image and modification-text paths before contextualized fusion; its target is matching target images rather than retrieving scenarios from a scenario database (Zhang et al., 27 Mar 2026). This supports a narrower contemporary meaning in which dual-path scenario retrieval is primarily a retrieval-and-fusion design for scenario understanding, reasoning, or mining.

2. Canonical architectural patterns

The main architectures differ in what each path retrieves, whether both paths run in parallel, and how their outputs are fused.

System	Path pair	Combination mechanism
DaPT	source-language path and English-translation path	sub-question graph fusion, bilingual retrieval-and-answer
Hybrid-IR	graph-based retrieval and dense retrieval	reciprocal rank fusion inside iterative retrieve–reason loop
“Why Braking?”	category-based path and embedding-based path	route by known category or cosine retrieval
DTR	query-based retrieval and pseudo-context-based retrieval	union plus adaptive information selection
SPIKE	document-level retrieval and scenario-level retrieval	linear score $\lambda r_s + (1-\lambda) r_d$
RF-Mem	Familiarity path and Recollection path	mean-similarity and entropy gating
SMc2f	coarse VLM filtering and fine text–trajectory matching	prune then re-rank
DualResearch	breadth semantic graph and depth causal graph	entropy-gated log-linear fusion

Some systems run both paths in parallel and fuse them immediately. DaPT generates sub-question graphs in parallel for both the source-language query and its English translation counterpart, then merges them before bilingual retrieval (Wang et al., 19 Mar 2026). Hybrid-IR fans each sub-question into a graph retriever and a dense retriever, then fuses the rankings by reciprocal rank fusion (Wan et al., 24 Jun 2026). DualResearch similarly queries a breadth semantic graph and a depth causal graph and fuses answer distributions in log space via entropy gating (Shi et al., 10 Oct 2025).

Other systems use adaptive path selection rather than unconditional parallelism. RF-Mem computes a familiarity signal from mean similarity $\mu$ and entropy $H$ ; high familiarity leads to the direct top- $K$ Familiarity path, while low familiarity activates the Recollection path (Zhang et al., 10 Mar 2026). DTR first measures generation uncertainty $u$ , and only when $u>\tau$ does it invoke retrieval; once triggered, it uses both the original query and a generated pseudo-context as complementary retrieval paths (Chen et al., 7 Jan 2026). In autonomous-driving scenario mining, the “Why Braking?” framework uses a category-based path for known scenario types and an embedding-based path for unknown out-of-distribution scenarios (Wu et al., 17 Jul 2025).

A third pattern is sequential dual-path operation. SMc2f first performs CLIP-based coarse filtering over image windows and then applies a fine-grained text–trajectory matcher to re-rank candidate trajectories returned by an LLM-based mining stage (Chen et al., 17 Jan 2026). This suggests that “dual-path” does not require simultaneous evidence construction; it may also refer to a prune-then-refine pipeline.

3. Retrieval units, path construction, and scoring

In DaPT, the two paths are defined over multilingual question representations. Let $q_\ell$ be the original non-English question and $q_{en}$ its English translation via GPT-4-mini. The framework constructs two acyclic sub-question graphs,

$6$0

where nodes are single-hop sub-questions and directed edges encode dependency. DaPT then fuses semantically identical nodes by embedding each sub-question with BGE-m3, computing cosine similarity, and merging the best pair when similarity exceeds $6$1. After topological sorting, each node is solved sequentially; bilingual nodes retrieve in both languages in parallel, and passage ranking uses

$6$2

Top-$6$3 retrieval uses $6$4 for most methods and $6$5 for HippoRAG2, and bilingual candidates are reconciled by a lightweight Judge step followed by Select or ReGen (Wang et al., 19 Mar 2026).

Hybrid-IR instantiates dual-path retrieval over heterogeneous substrates. Offline, it extracts OpenIE-style triples $6$6, augments them with undirected synonymy edges when Contriever-MSMARCO cosine similarity exceeds $6$7, and builds a heterogeneous graph

$6$8

Entity–document provenance is stored in a binary matrix $6$9. Online, query mentions $\lambda r_s + (1-\lambda) r_d$ 0 are aligned to KG entities, a Graph Foundation Model (GFM-RAG–8M) performs message passing, and document relevance is aggregated as

$\lambda r_s + (1-\lambda) r_d$ 1

In parallel, MedCPT embeds documents and the query to compute

$\lambda r_s + (1-\lambda) r_d$ 2

The two rankings are subsequently fused (Wan et al., 24 Jun 2026).

In autonomous-driving braking analysis, the retrieval unit is a time window $\lambda r_s + (1-\lambda) r_d$ 3 over a multivariate signal sequence $\lambda r_s + (1-\lambda) r_d$ 4. The category-based path uses predicates $\lambda r_s + (1-\lambda) r_d$ 5 or, in practice, an LLM over duration-based summaries of rule tags to assign $\lambda r_s + (1-\lambda) r_d$ 6. The embedding-based path learns $\lambda r_s + (1-\lambda) r_d$ 7 with a triplet loss

$\lambda r_s + (1-\lambda) r_d$ 8

where the margin is typically set to $\lambda r_s + (1-\lambda) r_d$ 9. Retrieval uses cosine similarity and an OOD threshold $\mu$ 0, with an example value $\mu$ 1 (Wu et al., 17 Jul 2025).

SPIKE uses a dual-vector-store formulation. Each document $\mu$ 2 is paired with scenarios $\mu$ 3, where each scenario is a short natural-language reasoning unit structured into Main Topic, Key Aspects, Information Need, and Explanation. Inference retrieves both top- $\mu$ 4 documents and top- $\mu$ 5 scenarios, then combines document-level and scenario-level signals through

$\mu$ 6

with $\mu$ 7 in the reported setting (Lee et al., 29 Mar 2025).

DTR and RF-Mem define the two paths in terms of query reformulation and cognitive-style retrieval. DTR embeds the raw query $\mu$ 8 for one path and a generated pseudo-context $\mu$ 9 for the other. RF-Mem begins with probe retrieval, computes $H$ 0 and $H$ 1, and if uncertainty is high enters a Recollection path that clusters retrieved candidates and updates the query embedding by

$H$ 2

This makes the second path a controlled multi-round expansion rather than a second index lookup (Chen et al., 7 Jan 2026, Zhang et al., 10 Mar 2026).

In robotic autonomy, SMc2f uses different modalities across its two paths. The coarse path encodes sampled RGB frames and text with CLIP ViT-L/14, scores sliding windows of length $H$ 3 and stride $H$ 4, selects top- $H$ 5 windows, and discards trajectories outside the merged coarse interval. The fine path projects trajectories with PatchTST and text with a CLIP-based text encoder into a shared $H$ 6-dimensional space and re-ranks LLM-produced candidates by dot-product similarity (Chen et al., 17 Jan 2026). DualResearch likewise separates background and procedural evidence into a breadth semantic graph $H$ 7 and a depth causal graph $H$ 8, turning retrieved paths into answer distributions rather than directly retrieving documents (Shi et al., 10 Oct 2025).

4. Fusion, gating, and iterative control

The central technical problem in dual-path systems is not merely constructing two paths, but deciding when and how to combine them. DaPT resolves path divergence by node fusion. After finding

$H$ 9

it redirects edges from the English node to the source-language node when similarity exceeds $K$ 0, marks the retained node as bilingual, and repeats until no pair exceeds the threshold. At answer time, a bilingual node yields $K$ 1 and $K$ 2; if Judge returns True, Select prefers the $K$ 3-answer, otherwise ReGen uses the entire bilingual reasoning path as context (Wang et al., 19 Mar 2026).

Hybrid-IR addresses score heterogeneity by rank-level fusion. It ranks documents separately by graph relevance and dense similarity, then computes

$K$ 4

with $K$ 5 described as a small smoothing constant and exemplified by $K$ 6. The fused retrieval is embedded in an iterative retrieve–reason loop over rounds $K$ 7, where the history state $K$ 8 stores the original question and all collected sub-question/sub-answer pairs. The loop stops when $K$ 9 or when the LLM signals “No further sub-questions needed,” and the reported architecture layers the retrieve–reason step three times $u$ 0 (Wan et al., 24 Jun 2026).

DTR uses uncertainty as a trigger rather than a fusion weight. For an autoregressively generated answer $u$ 1, the uncertainty score is

$u$ 2

If $u$ 3, the system returns the parametric answer directly; otherwise it retrieves along both query and pseudo-context paths. Candidate documents $u$ 4 are re-ranked by the adaptive information selection score

$u$ 5

where $u$ 6 and $u$ 7 are the path-specific inner products. The paper sweeps $u$ 8 over $u$ 9, with $u>\tau$ 0 as a default (Chen et al., 7 Jan 2026).

RF-Mem also uses a gate, but the gate is based on retrieval uncertainty rather than generation uncertainty. The strategy is

$u>\tau$ 1

and when $u>\tau$ 2, entropy $u>\tau$ 3 decides between the two. The Familiarity path returns the top- $u>\tau$ 4 memory items immediately; the Recollection path performs B-way beam search with fan-out $u>\tau$ 5 over $u>\tau$ 6 rounds, with typical settings $u>\tau$ 7, $u>\tau$ 8, $u>\tau$ 9 (Zhang et al., 10 Mar 2026).

DualResearch makes uncertainty itself a fusion signal. It converts breadth-path and depth-path evidence into channel-specific answer distributions $q_\ell$ 0 and $q_\ell$ 1, computes Shannon entropies $q_\ell$ 2 and $q_\ell$ 3, then defines

$q_\ell$ 4

The fused answer score is

$q_\ell$ 5

followed by a global calibration term that penalizes joint uncertainty (Shi et al., 10 Oct 2025). Across these systems, fusion may occur at the level of nodes, document ranks, query routing, or answer distributions, but it is consistently query-dependent.

5. Application domains and empirical behavior

In multilingual multi-hop QA, dual-path retrieval is used to compensate for the imbalance between English-centered LLM reasoning and multilingual evidence fidelity. DaPT reports that advanced RAG systems suffer from a significant performance imbalance in multilingual scenarios, and on the most challenging MuSiQue benchmark it achieves a relative improvement of $q_\ell$ 6 in average EM score over the strongest baseline while yielding more accurate and concise answers than baselines (Wang et al., 19 Mar 2026).

In complex medical QA, Hybrid-IR uses graph retrieval to capture global, multi-hop entity connections and dense retrieval to retain fine-grained semantic matching. On MedQA, MedMCQA, and MMLU-Med, using a strong $q_\ell$ 7B Llama-3 backbone, Hybrid-IR achieves $q_\ell$ 8 average accuracy versus $q_\ell$ 9 for the prior iterative baseline i-MedRAG; with GPT-4o-mini, it improves from $q_{en}$ 0 to $q_{en}$ 1. Ablation results state that removing either the graph path, the vector path, or the iterative loop all hurt performance, indicating that the three components are jointly essential (Wan et al., 24 Jun 2026).

In open-domain QA and dense retrieval, the main effect of dual-path design is better handling of sparse or ambiguous queries and reasoning-intensive implicit relevance. DTR reports average EM/F1 of $q_{en}$ 2 over five open-domain QA datasets with Qwen2.5-7B-Instruct and bge, compared with $q_{en}$ 3 for Standard RAG. On HotpotQA, recall@3 improves from $q_{en}$ 4 to $q_{en}$ 5 with bge and from $q_{en}$ 6 to $q_{en}$ 7 with e5 under dual-path retrieval. SPIKE reports nDCG@10 gains on BRIGHT from $q_{en}$ 8 to $q_{en}$ 9 depending on the dense retriever, and still adds $6$00–$6$01 on top of GPT-4 chain-of-thought-augmented queries (Chen et al., 7 Jan 2026, Lee et al., 29 Mar 2025).

In autonomous-driving scenario mining, the dual-path design separates known and unknown scenario retrieval. The “Why Braking?” framework annotates $6$02 driving logs into $6$03 scenarios across $6$04 classes. For known in-distribution scenario retrieval, overall rule-only performance is $6$05, $6$06, $6$07, while the dual-path, LLM-classification method reaches $6$08, $6$09, $6$10. Disabling Path A reduces F1 from $6$11 to $6$12, and disabling Path B makes OOD retrieval impossible, with zero recall. In robotic autonomy, SMc2f reports validation-set gains over RefAV from $6$13 to $6$14 on HOTA-T, from $6$15 to $6$16 on HOTA, from $6$17 to $6$18 on Timestamp-F1, and from $6$19 to $6$20 on Log-F1, while reducing per-query–log runtime from $6$21 to $6$22 on RTX 4090 (Wu et al., 17 Jul 2025, Chen et al., 17 Jan 2026).

In personalization and answer reconstruction, dual-path retrieval supports adaptive depth. RF-Mem improves PersonaMem generation Accuracy by $6$23–$6$24 absolute over Dense Retrieval and Full-Context across corpus scales, with $6$25–$6$26 retrieval times and only $6$27–$6$28 input tokens versus $6$29–$6$30 in Full-Context. On PersonaBench it achieves Recall@10 up to $6$31, and on LongMemEval it matches or exceeds Recall@50 $6$32 with $6$33 latency. DualResearch, using execution logs from InternAgent, improves accuracy by $6$34 on HLE and $6$35 on GPQA; with o4-mini on GPQA Diamond, $6$36 rises to $6$37 (Zhang et al., 10 Mar 2026, Shi et al., 10 Oct 2025).

Dual-path retrieval typically introduces additional control parameters, index structures, or iterative compute. Hybrid-IR reports that performance gains saturate around $6$38, after which marginal benefit diminishes while latency increases roughly linearly. SPIKE states that the dual-vector-store design roughly doubles index size and requires two ANN lookups, with the extra retrieval adding $6$39. RF-Mem notes that typical settings $6$40, $6$41, $6$42 keep Recollection cost within a $6$43 of Familiarity, and DTR uses uncertainty gating to reduce retrieval calls to $6$44 at $6$45 for the $6$46B model and $6$47 at the same threshold for the $6$48B model (Wan et al., 24 Jun 2026, Lee et al., 29 Mar 2025, Zhang et al., 10 Mar 2026, Chen et al., 7 Jan 2026).

The error profile also shifts with path design. In braking-scenario mining, anticipatory, low-risk braking without TrajOverlapTag may be missed by KeyIdent, and lacking HD maps can cause the LLM to confuse left_oppo with obj_crossing. In multilingual QA, the DaPT formulation itself is motivated by the observation that overreliance on English semantic understanding diminishes effectiveness in multilingual scenarios, so its dual-path design treats source-language fidelity and English reasoning robustness as separate assets to be reconciled (Wu et al., 17 Jul 2025, Wang et al., 19 Mar 2026).

A common misconception is that dual-path retrieval always means parallel retrieval over the same corpus with the same retrieval objective. The literature shows otherwise: RF-Mem switches between paths adaptively, SMc2f applies a coarse path before a fine path, DaPT retrieves over $6$49 and $6$50, and DualResearch fuses answer distributions from different graph substrates rather than simply concatenating documents (Zhang et al., 10 Mar 2026, Chen et al., 17 Jan 2026, Wang et al., 19 Mar 2026, Shi et al., 10 Oct 2025). Another misconception is that adding a second path automatically improves robustness. Saboori et al. report that, under node failures only, Crowds outperforms Dual-Path by $6$51, whereas under node failures plus rising traffic, Dual-Path regains a positive improvement ratio up to $6$52 (Saboori et al., 2012). This suggests that the value of a second path depends on the failure model, the fusion rule, and the operational cost of maintaining both paths.

Overall, dual-path scenario retrieval is best understood as a retrieval architecture for complementarity management: one path alone rarely captures all of structure, semantics, uncertainty, provenance, and domain specificity, so a second path is introduced to cover the missing regime. The recent literature indicates that the decisive design variable is not merely the presence of two paths, but how their heterogeneity is scored, routed, and fused under task-specific uncertainty (Wan et al., 24 Jun 2026, Chen et al., 7 Jan 2026)