Pool-Conditioned Query Resampling

• Pool-conditioned query resampling is a design paradigm where resampling, query construction, or answer generation is explicitly conditioned on an existing candidate pool to optimize performance.
• Techniques like the two-stage attention in PARCEL and candidate auditing in R³-SQL illustrate how pool-awareness preserves spatial grounding and improves execution accuracy.
• Applications span adaptive data analysis and synthetic query processing, demonstrating that leveraging available pool structure reduces aliasing, enhances generalization, and streamlines computation.

Searching arXiv for the cited papers to ground the article in current records. Pool-conditioned query resampling denotes a family of pool-aware procedures in which resampling, query construction, or answer generation is explicitly conditioned on an existing pool of candidates, tokens, queries, or data points. In the cited literature, the phrase spans several technically distinct settings: a two-stage attention mechanism in visual token compression, selective candidate-pool replacement in Text-to-SQL, continuously refreshed query sampling for search-quality measurement, fresh-subsample answering in adaptive data analysis, and conditional synthetic sampling for approximate group-by query processing (Kuzucu et al., 28 May 2026, Han et al., 28 Apr 2026, Astrakhantsev et al., 2022, Blanc, 2023, Zhang et al., 2021). This suggests that the term is not a single canonical algorithm but a recurring design pattern in which a downstream query or resampling step is constrained by structure already present in a pool.

1. Scope, terminology, and recurring abstractions

Across these works, the operative meaning of pool varies. In PARCEL, the pool consists of spatial pool tokens derived from raw ViT features $X_v$, and the query tokens are conditioned on those anchors before cross-attending to the original features. In R$^3$-SQL, the pool is a candidate set $S=\{s_1,\dots,s_n\}$ of SQL programs, and an audit agent decides whether that pool is sufficient or should be discarded and resampled. In continuous search-quality monitoring, the pool is a historical set $P_t$ of past user queries, from which a fixed-size judgment sample $S_t\subset P_t$ is repeatedly extracted. In adaptive data analysis, the pool is a held-out dataset $D\sim P^n$, and each adaptive query is answered on a fresh random subsample. In approximate query processing for group-by queries, the pool is the underlying database $D$, while resampling is mediated by a conditional generative model that produces synthetic tuples conditioned on a query label (Kuzucu et al., 28 May 2026, Han et al., 28 Apr 2026, Astrakhantsev et al., 2022, Blanc, 2023, Zhang et al., 2021).

The corresponding notion of conditioning is equally heterogeneous. It may be architectural, as in multi-head self-attention over $[P;Q_{\mathrm{IN}}]$; decision-theoretic, as in a binary audit function $f(S)\in\{0,1\}$; distributional, as in weights $w_{i,t}$ over a query population; or statistical, as in the restriction that each adaptive query sees only a random $^3$0-subset of the data pool. The common structure is that resampling is not unconditional: it depends on information extracted from the existing pool.

A common misconception is to read the phrase as synonymous with naive resampling or with unconditional query compression. The surveyed works do not support that interpretation. They instead use pool structure to control aliasing, preserve grounding, improve candidate recall, regulate label reuse, or guarantee generalization. The resulting procedures are therefore best understood as pool-aware resampling mechanisms rather than generic resampling heuristics.

2. Pool-Conditioned Query Resampling in PARCEL

In "PARCEL: Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding," Pool-Conditioned Query Resampling (PCQR) is the central two-stage attention block in a hybrid visual connector (Kuzucu et al., 28 May 2026). PARCEL splits the compression work into two pathways: a fixed, grid-aligned Pool-Anchor pathway that anchors low-frequency, geometric layout via spatial average pooling, and a Query pathway of learnable semantic-explorer tokens that recover complementary, high-frequency detail.

Formally, the raw ViT features are $^3$1 with $^3$2. Given a total budget $^3$3, PCQR splits $^3$4 into $^3$5 pool tokens and $^3$6 query tokens through budget-aware routing. Pool tokens are obtained by average-pooling: $^3$7 A fixed bank of learnable query embeddings $^3$8 is maintained, and nested dropout during training keeps only the first $^3$9 of them. PCQR then forms the joint sequence

$S=\{s_1,\dots,s_n\}$0

and applies a standard MSA block,

$S=\{s_1,\dots,s_n\}$1

The output is split into updated anchors $S=\{s_1,\dots,s_n\}$2 and pool-aware queries $S=\{s_1,\dots,s_n\}$3. These pool-aware queries then cross-attend to the uncompressed features: $S=\{s_1,\dots,s_n\}$4 The final compressed visual representation is the concatenation $S=\{s_1,\dots,s_n\}$5.

The algorithmic sketch specifies the budget routing rule

$S=\{s_1,\dots,s_n\}$6

with $S=\{s_1,\dots,s_n\}$7. Average pooling constructs $S=\{s_1,\dots,s_n\}$8, the learnable bank supplies $S=\{s_1,\dots,s_n\}$9, a joint self-attention block produces $P_t$0, and cross-attention to $P_t$1 yields $P_t$2. PARCEL uses the same LVLM pretraining and transfer-fine-tuning objectives as PaliGemma-2, without introducing auxiliary losses. Nested Dropout on $P_t$3 enforces the nested prefix structure, and conditioning $P_t$4 on $P_t$5 is described as implicitly encouraging complementary information in $P_t$6.

The architectural role of PCQR is framed as a division of labor. Pooling preserves explicit 2D structure and low frequencies; queries, after being primed by the pooled anchors, recover complementary detail. The paper states that forcing queries to see the pooled anchors first frees them from re-encoding low-frequency layout, thereby avoiding spectral aliasing of pooling, while preserving explicit spatial grounding and avoiding the non-locality of pure query resampling.

Empirically, PARCEL is evaluated across 27 benchmarks and is reported to consistently outperform both Matryoshka-Multimodal Models (M$P_t$7) and the Matryoshka Query Transformer (MQT) at all budgets $P_t$8, $P_t$9, and $S_t\subset P_t$0 tokens. Table 1 reports image mean retention at 256 tokens of M$S_t\subset P_t$1 $S_t\subset P_t$2, MQT $S_t\subset P_t$3, and PARCEL $S_t\subset P_t$4; at 64 tokens, PARCEL $S_t\subset P_t$5 versus MQT $S_t\subset P_t$6. On RefCOCO at 256 tokens, PARCEL achieves $S_t\subset P_t$7 versus MQT $S_t\subset P_t$8 $S_t\subset P_t$9. On ChartQA, PARCEL yields $D\sim P^n$0 and $D\sim P^n$1 absolute points over M$D\sim P^n$2 at 64 and 256 tokens, respectively. The efficiency curves in Figure 1 are described as exhibiting a strictly better accuracy-vs-compute Pareto frontier under aggressive compression while preserving the "train once, deploy anywhere" paradigm.

3. Agentic pool-conditioned resampling in Text-to-SQL

In "R$D\sim P^n$3-SQL: Ranking Reward and Resampling for Text-to-SQL," the pool-conditioned mechanism is not token compression but candidate-pool auditing and selective replacement (Han et al., 28 Apr 2026). Modern Text-to-SQL systems generate multiple candidate SQL queries and rank them, but the paper identifies two limitations: functionally equivalent SQL queries may be scored inconsistently despite identical execution results, and ranking cannot recover when the correct SQL is absent from the candidate pool.

R$D\sim P^n$4-SQL first samples an initial pool $D\sim P^n$5 of SQL candidates using an LLM with temperature $D\sim P^n$6 and $D\sim P^n$7. It then invokes a lightweight audit agent $D\sim P^n$8 to predict whether $D\sim P^n$9 contains at least one likely-correct SQL. The agent takes as input the user question $D$0, the database schema $D$1, each candidate $D$2, and a few-row execution preview $D$3. It outputs a binary decision $D$4, where $D$5 means “pool is sufficient” and $D$6 means “pool is likely missing a correct SQL.” The resampling rule is

$D$7

When resampling is triggered, a fresh pool of size $D$8 is drawn, each candidate is scored by the pointwise ranker $D$9, the top-$[P;Q_{\mathrm{IN}}]$0 candidates are retained, and the original pool is replaced.

The resampled or retained pool is then subjected to unified reward-based ranking. Candidates are grouped by execution result, and each group $[P;Q_{\mathrm{IN}}]$1 receives two signals. The pairwise preference signal is derived from cross-group comparisons using

$[P;Q_{\mathrm{IN}}]$2

with $[P;Q_{\mathrm{IN}}]$3, and the groupwise list score is

$[P;Q_{\mathrm{IN}}]$4

The pointwise-group utility signal is

$[P;Q_{\mathrm{IN}}]$5

where $[P;Q_{\mathrm{IN}}]$6 is the group size and $[P;Q_{\mathrm{IN}}]$7 is the reciprocal-rank of $[P;Q_{\mathrm{IN}}]$8 under $[P;Q_{\mathrm{IN}}]$9. Groups are sorted by the tuple $f(S)\in\{0,1\}$0 in descending lexicographic order, and the final SQL is

$f(S)\in\{0,1\}$1

from the top group $f(S)\in\{0,1\}$2. During training of the GRPO-based listwise model R$f(S)\in\{0,1\}$3-7B, each positive/negative pair $f(S)\in\{0,1\}$4 receives reward

$f(S)\in\{0,1\}$5

where $f(S)\in\{0,1\}$6 if the model ranks $f(S)\in\{0,1\}$7 above $f(S)\in\{0,1\}$8, $f(S)\in\{0,1\}$9 if the ranking remains correct under input-order swap, and $w_{i,t}$0.

The empirical effect of the resampling stage is reported directly. Without agentic resampling, the recall of correct SQL across five benchmarks averages $w_{i,t}$1; with resampling it rises to $w_{i,t}$2 $w_{i,t}$3. Combined with groupwise ranking, R$w_{i,t}$4-SQL achieves $w_{i,t}$5 execution accuracy on BIRD-dev and an average $w_{i,t}$6 over BIRD-dev, Spider, Spider-DK, EHR-SQL, and ScienceBenchmark. In ablations, naive always-resample gives only $w_{i,t}$7, union of original and resampled pools hurts performance, and selective replacement with $w_{i,t}$8 yields the full $w_{i,t}$9 gain. The audit agent is a small SFT-fine-tuned LLM, Qwen2.5-Coder-Instruct, and on BIRD-dev it attains $^3$00 precision and $^3$01 recall for “resample” decisions. A plausible implication is that the value of the mechanism lies less in unconditional diversification than in deciding when the existing pool is untrustworthy.

4. Stable and semi-stable resampling for continuously used query pools

In "Stable and Semi-stable Sampling Approaches for Continuously Used Samples," pool-conditioned query resampling is formulated as a repeated sampling problem over a historical pool $^3$02 of user queries (Astrakhantsev et al., 2022). At each period $^3$03, a fresh judgment sample $^3$04 of fixed size $^3$05 must balance three competing objectives: representativeness with respect to the current query-traffic distribution $^3$06, labeling-cost minimization through reuse of previously judged queries, and overfitting resistance through controlled refresh.

Using inclusion indicators $^3$07 with $^3$08, the formulation introduces

$^3$09

$^3$10

and

$^3$11

The unified objective is

$^3$12

subject to $^3$13 and $^3$14.

The paper does not solve this discrete optimization directly. Instead, Stable and Semi-stable methods are presented as heuristic solutions. Stable SRS reuses the same uniform keys $^3$15 for continuing queries and only generates fresh keys for newly appearing queries, then selects the top-$^3$16 keys. Semi-stable SRS fixes a desired monthly refresh fraction $^3$17, partitions $^3$18 into $^3$19 shards, and refreshes one shard per round. Stable WRS is based on Efraimidis-Spirakis weighted random sampling with keys

$^3$20

keeping the same base random $^3$21 for continuing queries across time. Semi-stable WRS maintains two hashing seeds $^3$22, a refresh accumulator $^3$23, and mixes the two hash-derived uniforms to achieve a controlled refresh schedule without persistent storage of per-query randomness.

The stated theoretical property is exact validity: Stable and Semi-stable variants preserve the exact inclusion-probability properties of standard SRS/WRS at each $^3$24. Representativeness is tied to $^3$25, the stability penalty is controlled by reusing keys, and labeling cost is precisely the count of newly included queries. Standard sampling-without-replacement variance results still apply, with $^3$26.

The empirical evaluation uses Microsoft Bing query logs from 2019–2020 for Image Results Page (IRP) and Search Engine Results Page (SERP), with sample size $^3$27 rolling monthly for 12 months. Mean consecutive-month overlap on IRP is reported as WRS $^3$28, Stable WRS $^3$29, Semi-stable WRS with $^3$30 $^3$31, and Semi-stable WRS with $^3$32 $^3$33. On SERP, WRS gives $^3$34, Stable WRS $^3$35, and Semi-stable WRS with $^3$36 $^3$37. Judgment load stabilizes after month 1 at

$^3$38

reaching $^3$39 for IRP with $^3$40 and $^3$41 with $^3$42. The paper’s practical guidance is correspondingly explicit: Stable WRS is recommended when labeling budget is extremely constrained and population changes slowly, while Semi-stable with $^3$43 is recommended when some resistance to overfitting is needed or labels age out after $^3$44 periods.

5. Fresh-subsample answering in adaptive data analysis

In "Subsampling Suffices for Adaptive Data Analysis," the relevant mechanism is answering each adaptively chosen query on a fresh random subsample of a fixed data pool $^3$45 (Blanc, 2023). The analyst issues queries $^3$46 adaptively; at step $^3$47, after observing $^3$48, the analyst chooses a function

$^3$49

with $^3$50, so each answer communicates at most $^3$51 bits. The held-out dataset is reused across queries, but each query only sees a random $^3$52-subset.

The mechanism is simple. For each $^3$53, receive $^3$54, draw a fresh random subsample $^3$55 of size $^3$56, and output $^3$57. Without replacement, $^3$58 is chosen uniformly among all $^3$59-subsets of $^3$60, denoted $^3$61. The paper’s main claim is that the randomness of subsampling alone is sufficient to guarantee generalization for adaptive sequences of few-bit queries.

The theoretical route has four steps. First, subsampling queries reveal little information through Average Leave-Many-Out KL stability. Second, this stability implies a mutual-information bound: $^3$62 Choosing $^3$63 gives

$^3$64

Third, mutual information implies generalization via information-theoretic transfer. Fourth, combining these ingredients yields the informal bound

$^3$65

and by Markov one obtains a constant-probability bound of order

$^3$66

The paper states informally that if $^3$67, then an adaptive sequence of $^3$68 subsampling queries each of width $^3$69 and range size $^3$70 will all have small bias.

Two applications are given. For statistical queries $^3$71, the mechanism draws $^3$72, samples Bernoulli votes $^3$73, and returns $^3$74. If $^3$75 and

$^3$76

then with probability at least $^3$77, for all $^3$78,

$^3$79

For approximate-median queries, the mechanism partitions $^3$80 into $^3$81 disjoint blocks, obtains one subsample-vote per block, and performs binary search over $^3$82 in $^3$83 steps. The framing differs sharply from PARCEL or R$^3$84-SQL, but the shared structural feature is again visible: adaptive queries are constrained to interact only with randomized views of a fixed pool, and the paper argues that this is enough to control generalization without external noise or black-box differential privacy.

6. Conditional generative resampling for approximate group-by queries

In "Approximate Query Processing for Group-By Queries based on Conditional Generative Models," the relevant resampling mechanism is synthetic tuple generation conditioned on a query label $^3$85 (Zhang et al., 2021). The task is to approximate

$^3$86

for a large table $^3$87, where exact execution is I/O- and CPU-intensive when $^3$88 is large. The stated challenge is to achieve both accuracy and efficiency. Uniform sampling under-samples rare groups, stratified sampling requires foreknowledge of the aggregates or predicates, and online sampling incurs large I/O and latency at query time.

The proposed solution is to train a compact conditional generative model $^3$89 that approximates the true conditional density

$^3$90

where $^3$91 encodes the grouping label and, for predicates, a bucket label. At query time, $^3$92 can produce an arbitrary number of synthetic tuples $^3$93 without accessing $^3$94. The model is a Conditional Wasserstein GAN. The generator $^3$95 takes latent $^3$96 concatenated with $^3$97, uses fully connected layers $^3$98 with ReLU, and outputs $^3$99. The critic $S=\{s_1,\dots,s_n\}$00 takes $S=\{s_1,\dots,s_n\}$01 concatenated with $S=\{s_1,\dots,s_n\}$02, uses the same layer sizes and activations, and is trained with the Wasserstein GAN objective

$S=\{s_1,\dots,s_n\}$03

where

$S=\{s_1,\dots,s_n\}$04

At query time, the sample-generation procedure determines target counts $S=\{s_1,\dots,s_n\}$05 for each group $S=\{s_1,\dots,s_n\}$06, draws $S=\{s_1,\dots,s_n\}$07, sets $S=\{s_1,\dots,s_n\}$08 to the one-hot encoding of the relevant group or bucket, and emits

$S=\{s_1,\dots,s_n\}$09

These synthetic samples can be combined with real samples in a stratified estimator. For the sum $S=\{s_1,\dots,s_n\}$10, the estimator is

$S=\{s_1,\dots,s_n\}$11

and

$S=\{s_1,\dots,s_n\}$12

If $S=\{s_1,\dots,s_n\}$13 exactly reproduces $S=\{s_1,\dots,s_n\}$14, then $S=\{s_1,\dots,s_n\}$15 is unbiased. Under the assumption

$S=\{s_1,\dots,s_n\}$16

and $S=\{s_1,\dots,s_n\}$17 being $S=\{s_1,\dots,s_n\}$18-Lipschitz, the bias is at most $S=\{s_1,\dots,s_n\}$19, and the total mean-squared error satisfies

$S=\{s_1,\dots,s_n\}$20

The empirical evaluation uses ROAD and GAUSS, with baselines including uniform sampling, CVOPT stratification on real $S=\{s_1,\dots,s_n\}$21, online aggregation on real $S=\{s_1,\dots,s_n\}$22, and DBEst. The reported results include the following. On ROAD, with $S=\{s_1,\dots,s_n\}$23 total samples, CVOPT+$S=\{s_1,\dots,s_n\}$24 yields substantially narrower confidence intervals and lower ARE than random sampling. On GAUSS, as synthetic sample size grows from $S=\{s_1,\dots,s_n\}$25, CI width shrinks $S=\{s_1,\dots,s_n\}$26 and relative-CI/mean falls linear in $S=\{s_1,\dots,s_n\}$27. For selectivities $S=\{s_1,\dots,s_n\}$28 on ROAD, $S=\{s_1,\dots,s_n\}$29+predicate-bucketing is $S=\{s_1,\dots,s_n\}$30 more accurate than random and outperforms DBEst; at $S=\{s_1,\dots,s_n\}$31, $S=\{s_1,\dots,s_n\}$32 latency is $S=\{s_1,\dots,s_n\}$33 versus DBEst $S=\{s_1,\dots,s_n\}$34. Generating $S=\{s_1,\dots,s_n\}$35 samples from $S=\{s_1,\dots,s_n\}$36 is $S=\{s_1,\dots,s_n\}$37 time $S=\{s_1,\dots,s_n\}$38, whereas I/O sampling from $S=\{s_1,\dots,s_n\}$39 grows linearly with $S=\{s_1,\dots,s_n\}$40 and is reported as $S=\{s_1,\dots,s_n\}$41 on $S=\{s_1,\dots,s_n\}$42 tuples. The CWGAN footprint is approximately $S=\{s_1,\dots,s_n\}$43, compared with DBEst per-group models of $S=\{s_1,\dots,s_n\}$44 to $S=\{s_1,\dots,s_n\}$45. In this usage, pool-conditioned resampling does not mean refreshing a stored sample; it means replacing direct access to the pool with conditional synthetic sampling that is query-specific.

7. Comparative perspective

The cited works collectively show that pool-conditioned query resampling is a cross-domain design pattern rather than a single methodology. In PARCEL, conditioning is architectural and serves representational separation: pool anchors preserve explicit 2D structure and low frequencies, while conditioned queries recover complementary detail. In R$S=\{s_1,\dots,s_n\}$46-SQL, conditioning is decision-theoretic: an audit agent predicts whether the candidate pool is sufficient, and resampling is selective rather than unconditional. In continuous search monitoring, conditioning regulates the trade-off among representativeness, label reuse, and overfitting resistance. In adaptive data analysis, conditioning is statistical: every adaptive query is restricted to a fresh randomized projection of the same data pool. In approximate group-by query processing, conditioning is generative and allows arbitrary query-specific synthetic sample sizes without touching the base table.

Several contrasts follow directly. First, the object being resampled differs: visual queries, SQL candidates, user queries, data subsets, or synthetic tuples. Second, the objective differs: preserving spatial grounding, improving execution accuracy, minimizing labeling cost, guaranteeing generalization, or reducing query latency. Third, the notion of correctness differs: Pareto-frontier retention, execution accuracy and recall, exact inclusion-probability validity, transcript generalization, or unbiasedness and MSE control. What unifies them is that the pool is not merely a source of examples; it is an active conditioning signal.

This broader view also clarifies the main misconception surrounding the term. Pool-conditioned query resampling is not inherently about query compression, not inherently about dataset subsampling, and not inherently about label reuse. Those are domain-specific instantiations. The technically stable core is that resampling or query formation is conditioned on an already available pool, and the conditioning is designed to improve a concrete objective that naive resampling does not directly optimize.