Single-Query Paradigm Overview

• Single-Query Paradigm is a design approach that compresses multi-step processes into a single query-bearing unit for efficient computation.
• It is applied in diverse fields including medical segmentation, perception, retrieval systems, optimization, and quantum computing, often reducing complexity.
• Studies indicate its effectiveness is context-dependent, prompting alternatives like multi-query or iterative planning when single-query assumptions fall short.

Searching arXiv for papers on “single-query paradigm” and closely related formulations across domains. The single-query paradigm denotes a family of formulations in which a single query, prompt, embedding, access, or oracle interaction is treated as the primary computational unit. Across recent work, the term appears in several technically distinct senses: a single prompt or query initializer that conditions downstream prediction; a single-shot retrieval call or single embedding used to search tools or documents; a single unified query representation reused across tasks; a single-access algorithm for ranked answers; and a single exact quantum or oracle query in complexity-theoretic settings (Qu et al., 3 Nov 2025, Fang et al., 12 Jan 2026, Hu et al., 9 Jan 2026, Guoliang et al., 2021).

1. Terminological scope and recurring abstractions

The phrase does not name one invariant architecture. It names a recurrent design choice: compress control, conditioning, or verification into one query-bearing object, then either exploit that simplification or expose its limits.

Domain	Single-query unit	Operational meaning
3D medical segmentation	Single 3D point prompt	Seed exemplar expanded into multiple instance queries
Tool/document retrieval	One textual query or embedding	One retrieval call without explicit planning
Human-centric perception / detection	One unified query per instance or one unified query type	Same query reused across tasks or decoder modeling
Database / proof / quantum models	One access, one oracle query, or one quantum query	Ranked-answer access, one-challenge verification, or exact query complexity

In multi-task human-centric perception, HQNet defines a “Human Query” as the refined content query $q_i^{(L)} \in \mathbb{R}^d$ that represents one person instance and is reused for detection, segmentation, 2D keypoints, 3D mesh, gender, and age prediction (Jin et al., 2023). In complex query answering over knowledge graphs, SQE replaces operator-wise neural execution with a single sequence encoder applied to a linearized computational DAG, and uses the first hidden state as the query embedding for answer retrieval (Bai et al., 2023). In flexible object detection, DS-Det uses “Single-Query” to denote a unified query type derived directly from encoder tokens rather than separate content and positional queries (Cao et al., 26 Jul 2025). In database theory, “single access” is the single-access counterpart of direct access: compute the $k$-th ranked answer once, without a reusable index (Hu et al., 9 Jan 2026).

A common misconception is that “single-query” always means one query for the entire input or scene. The literature is more specific. In HQNet it means one query per human instance (Jin et al., 2023). In DS-Det it means one unified decoder query type (Cao et al., 26 Jul 2025). In SQE it means one encoder for the entire logical query (Bai et al., 2023). In database theory it means one-shot access to one rank position (Hu et al., 9 Jan 2026).

2. Unified query representations in perception and structured prediction

In medical image segmentation, MIQ-SAM3D explicitly departs from the standard single-point-to-single-object paradigm. A single 3D point $p=(d,h,w)$ is treated as a seed exemplar, from which the Prompt-Conditioned Instance Query Generator produces a set of instance queries: $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$ These queries are refined by the Competitive Query Refinement Decoder through inter-query self-attention and cross-attention to encoder features, so that a single clicked lesion can trigger segmentation of all semantically similar lesions in the volume (Qu et al., 3 Nov 2025). The model shifts the workflow from

$$\text{single point prompt} \Rightarrow \text{single seed prototype} \Rightarrow \text{multiple instance queries} \Rightarrow \text{multiple masks}.$$

On LiTS17, the full model reached Dice 60.47% and NSD 74.61%; removing PC-IQG + CQRD reduced Dice to 56.72% and removed the multi-instance behavior (Qu et al., 3 Nov 2025).

HQNet applies a different variant of the same idea. Each ground-truth person is assigned to one query by bipartite matching, and that same Human Query is decoded by lightweight task-specific heads for boxes, masks, keypoints, mesh, gender, and age (Jin et al., 2023). The architectural claim is not that there is one query for the image, but that there is one shared, task-agnostic query representation per person instance. On COCO-UniHuman, HQNet with ViT-L reports detection AP 78.0, segmentation AP 68.6, pose AP 75.3, gender AP 58.0, and age AP 58.0 (Jin et al., 2023).

DS-Det reformulates the DETR decoder around a unified Single-Query type. It removes the CQ/PQ split, generates flexible queries from encoder tokens through FLET, then disentangles localization and deduplication in the decoder: the Box Locating Part uses cross-attention with one-to-many matching, while the Deduplication Part uses cross-attention followed by self-attention under one-to-one matching (Cao et al., 26 Jul 2025). This is motivated by two identified pathologies: Recurrent Opposing inTeractions between self-attention and cross-attention, and “query ambiguity” when shared decoder weights are trained under both one-to-one and one-to-many assignments (Cao et al., 26 Jul 2025). With ResNet-50 for 12 epochs on COCO2017, DS-Det reports AP 50.5 versus 49.0 for DINO, and AP$_S$ 34.3 versus 32.0 (Cao et al., 26 Jul 2025).

A broader extension appears in end-to-end autonomous driving. SparseAD proposes a sparse query-centric paradigm in which sparse queries “completely represent the whole driving scenario across space, time and tasks without any dense BEV representation” (Zhang et al., 2024). This is not labeled a single-query system, but it pushes the same compression logic toward query-only scene state: perception, tracking, mapping, prediction, and planning all operate on learned query sets rather than a dense intermediate grid.

3. Single-shot retrieval as a bottleneck

In tool retrieval for LLM agents, the single-query paradigm is identified with single-shot dense retrieval: one user instruction, one query embedding, one retrieval call, one top-$k$ ranking (Fang et al., 12 Jan 2026). TOOLQP argues that this is structurally mismatched to large, dynamic tool libraries because user goals are high-level while tool documentation is low-level, because a single fixed-size embedding must represent compositional multi-tool intent, and because single-shot retrieval cannot inspect retrieved tools and refine subsequent queries (Fang et al., 12 Jan 2026). TOOLQP therefore recasts retrieval as iterative query planning, generating a plan $\mathcal{P}$, decomposing into sub-tasks $\{s_n\}$, and interacting with the retriever through a trajectory $\tau=\{(q_t,O_t)\}_{t=1}^T$ (Fang et al., 12 Jan 2026). On ToolRet, the full model reports zero-shot macro-average $k$0 and $k$1, compared with 29.8 and 27.0 for the base retriever (Fang et al., 12 Jan 2026).

ExpandSearch makes an analogous claim for search-augmented question answering. It characterizes prior RL-trained search agents as following “one query → one embedding → one retrieval list → long context → monolithic reasoning,” which degrades recall on multi-hop questions and produces noisy contexts (Zhao et al., 11 Oct 2025). Its response is an expand-then-squeeze loop: at each search turn the policy emits multiple queries $k$2, retrieves top-$k$3 passages for each, and passes the union to a frozen squeezer model that returns a compressed summary (Zhao et al., 11 Oct 2025). Across seven benchmarks, ExpandSearch reports an average Exact Match improvement of 4.4 percentage points over state-of-the-art baselines, and increasing the number of queries from $k$4 to $k$5 yields about 6.7% average EM gain (Zhao et al., 11 Oct 2025).

DocReRank inverts the usual query-centric mining pipeline. Instead of retrieving hard negative pages per query, it generates hard negative queries per page, creating a per-page training instance

$k$6

The paper explicitly describes this as a single-page or single-instance contrastive setup: given one page, distinguish the one positive query from several very similar but unanswerable queries (Wasserman et al., 28 May 2025). On ViDoReV2, DocReRank-Base reaches average NDCG@5 67.2 versus 64.4 for MonoQwen and 58.6 for ColQwen retrieval; on Real-MM-RAG it reaches 85.6 versus 80.9 for MonoQwen and 73.8 for ColQwen (Wasserman et al., 28 May 2025). This suggests that the single-query assumption becomes especially brittle when relevance depends on subtle paraphrases, latent dependencies, or compositional evidence aggregation.

4. Formal algorithmic regimes

In zeroth-order optimization, the single-query paradigm is not merely a modeling preference; it can be an optimal budget allocation rule. The standard single-query estimator is

$k$7

Under a fixed query budget $k$8, the paper analyzes two multi-query aggregation methods. For ZO-Avg, it proves that using more than one query per iteration is always query-inefficient across strongly convex, convex, non-convex, and stochastic settings; the optimal choice is $k$9. For ZO-Align, the conclusion reverses: more queries per iteration improve estimation, and full-subspace estimation $p=(d,h,w)$0 is optimal in the strongly convex and stochastic convex analyses (Lin et al., 19 Sep 2025). The paper therefore argues that single-query optimality is not universal but aggregation-method dependent.

In quantum query complexity, the single-query regime concerns exact computation with one oracle call. For partial Boolean functions, the paper proves that every function with $p=(d,h,w)$1 can be transformed to a simple partial Boolean function whose polynomial degree is one, and that for partial Boolean functions up to four bits there are only 10 non-trivial partial Boolean functions with a single quantum query (Guoliang et al., 2021). The structural consequence is that, up to the paper’s equivalence notion, the effective single-query class is governed by degree-1 representations rather than the looser general upper bound $p=(d,h,w)$2 (Guoliang et al., 2021).

A cryptographic analogue appears in proof-of-sequential-work. Prior PoSWs typically require $p=(d,h,w)$3 oracle queries to verify each challenge because commitment verification itself traverses a Merkle-like path (Sur, 2022). The proposed construction replaces that layer with an algebraic commitment mechanism, so that for each challenge the verifier uses one random-oracle query to recompute the challenged label and then checks

$p=(d,h,w)$4

The verifier thus needs only a single oracle query per challenge, while the prover still requires $p=(d,h,w)$5-sequential work under the random-oracle-based sequentiality argument (Sur, 2022).

5. Fixed-query evaluation in data systems

Database theory uses the term in a more literal way. Direct access assumes a fixed query $p=(d,h,w)$6, a fixed database instance $p=(d,h,w)$7, a total order $p=(d,h,w)$8 over answers, and an index $p=(d,h,w)$9; after preprocessing, the system returns the $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$0-th answer $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$1 without enumerating $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$2 (Hu et al., 9 Jan 2026). Its single-access counterpart solves the same task without reusable preprocessing, computing one rank position from scratch (Hu et al., 9 Jan 2026). For tractable conjunctive queries and lexicographic or sum orders, the implemented bounds are $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$3 preprocessing with $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$4 access for direct access, versus $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$5 time with no preprocessing for single access (Hu et al., 9 Jan 2026). In practice, the reported ratio between direct-access time for one access and single-access time for one access lies in $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$6, implying a very small practical break-even point (Hu et al., 9 Jan 2026).

The same fixed-query perspective is generalized in multi-branch analytics. “Querying Everything Everywhere All at Once” argues that agentic lakehouses are no longer adequately modeled as single-snapshot systems; many branches coexist, and consumers may need answers without naming a branch (Tagliabue, 11 Mar 2026). The paper formalizes per-branch classical evaluation and then interprets results under subvaluational truth conditions: $$Q_{\text{inst}} = \mathcal{G}_\theta(v_{\text{seed}}) \in \mathbb{R}^{N\times C}.$$7 In the prototype, number, boolean, and list-valued results are each rendered across branches rather than at one snapshot (Tagliabue, 11 Mar 2026). For booleans, the system distinguishes consensus truth, consensus falsity, and “mixed” outcomes when some branches satisfy the predicate and some do not (Tagliabue, 11 Mar 2026). This is still a single-query paradigm, but the query ranges over a multi-branch state rather than a single version.

6. Limits, misconceptions, and open directions

The literature also makes clear that the single-query paradigm is often a simplifying assumption rather than a stable optimum. MIQ-SAM3D assumes that all target lesions are semantically similar to the clicked exemplar, and the paper notes that heterogeneous lesion types, extremely small lesions, and generalization beyond tumor segmentation in CT remain open (Qu et al., 3 Nov 2025). TOOLQP explicitly remarks that iterative planning may be unnecessary for very small toolsets, implying that single-shot retrieval can remain adequate in simple settings (Fang et al., 12 Jan 2026). DocReRank depends on strong LLM/VLM generation and verification, and it identifies query variability, verification errors, and retriever dependence as limitations (Wasserman et al., 28 May 2025). Direct access is only available for structurally compatible query/order classes, and integration into a DBMS optimizer remains open (Hu et al., 9 Jan 2026). In ZO optimization, single-query optimality holds for ZO-Avg but fails for ZO-Align (Lin et al., 19 Sep 2025).

Several recurrent misconceptions are corrected by these results. Single-query does not always mean one output object, because MIQ-SAM3D uses a single point prompt to initialize multiple instance queries (Qu et al., 3 Nov 2025). It does not always mean one query for the whole image, because HQNet uses one query per person instance (Jin et al., 2023). It does not always mean one retrieval episode, because single-page training in DocReRank is organized around one page with multiple positive and negative queries (Wasserman et al., 28 May 2025). It does not always mean a universal efficiency optimum, because both TOOLQP and ExpandSearch show settings in which iterative or parallel query generation materially improves performance (Fang et al., 12 Jan 2026, Zhao et al., 11 Oct 2025).

Open directions are correspondingly diverse. MIQ-SAM3D suggests multiple points or higher-level prompts such as scribbles or 3D boxes (Qu et al., 3 Nov 2025). DocReRank points to stronger generation and verification circuitry, multi-page instances, and hardness calibration (Wasserman et al., 28 May 2025). ExpandSearch motivates more structured or hierarchical query planning and adaptive query budgets (Zhao et al., 11 Oct 2025). Database direct access raises parallelization, dynamic maintenance, and broader tractable classes (Hu et al., 9 Jan 2026). In query complexity, extending the clarified one-query structure to richer multi-query regimes remains open in both zeroth-order optimization and quantum-query models (Lin et al., 19 Sep 2025, Guoliang et al., 2021).

Taken together, these works show that the single-query paradigm is best understood as a controlled compression principle: one query-bearing object is made to stand in for a richer search, reasoning, or verification process. Where the compression aligns with the task structure, it yields architectural simplicity, efficient supervision, or optimal query complexity. Where it does not, recent work replaces single-query operation with multi-instance expansion, multi-step planning, or branch-aware semantics rather than treating the original assumption as fundamental.