Quality-Controllable Retrieval

• Quality-controllable retrieval is a framework where retrieval systems are steered by explicit control signals to meet specified quality profiles such as precision, diversity, and relevance.
• It employs methods like interpolation between retrieval objectives, conditional query generation, and post-processing adjustments to dynamically balance competing performance metrics.
• Applications span search, recommendation, and retrieval-augmented generation, with measurable impacts on metrics including recall, precision, and watch time.

Searching arXiv for papers on quality-controllable retrieval and related controllable retrieval frameworks. Quality-controllable retrieval denotes retrieval systems in which retrieval behavior is explicitly steered toward a specified target quality profile, typically through a control function, a conditional variable, or an inference-time parameter. In the controllable learning formalization for information retrieval, the requirement space is written as $$\mathcal{T}=\{\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}},\mathbf{s}_{\mathrm{tgt}}\}\in\Gamma$$, and a control function $h$ maps a base learner $f$ and the requirement description and context to a controlled learner $f_{\mathcal{T}}$, with the crucial property that new requirements can be handled without retraining the base model (Shen et al., 2024). Within retrieval, “quality” is not a single universal metric: the target may be precision, diversity, recall, retrievability, watch time, correctness, efficiency, security, maintainability, relevance, aesthetics, factual consistency, or an explicit balance between accuracy and cost (Shen et al., 2024, Penha et al., 2023, Liu et al., 2024, Geng et al., 31 May 2025, Lu et al., 24 Feb 2026, Su et al., 17 Feb 2025).

1. Formalization and conceptual scope

The survey formulation of controllable learning places quality-controllable retrieval within a broader IR framework. Let $\mathcal{X}$ denote inputs such as queries or user profiles, let $\mathcal{Y}$ denote outputs such as ranked lists, and let $\Gamma$ denote task requirements. The controlled learner is written as

$$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$

where $\mathbf{s}_{\mathrm{desc}}$ is the machine-readable description of the target, $\mathbf{s}_{\mathrm{ctx}}$ is auxiliary context, and $h$0 is the desired numeric or semantic target (Shen et al., 2024). In this formulation, quality is simply one dimension of $h$1; for example, it may be “precision@10 $h$2,” or a weighted balance between precision and diversity expressed by a control weight $h$3 (Shen et al., 2024).

The literature represented by recent systems shows that “quality” is instantiated in multiple non-equivalent ways. In controllable video-to-music retrieval, it is the trade-off between broad audiovisual correspondence and label-driven retrieval specificity, controlled by the inference-time mixing weight $h$4 in

$h$5

(Stewart et al., 2024). In dense search, it may mean improving long-tail content discovery by controlling the narrow-versus-broad intent of generated queries through $h$6, where $h$7 (Penha et al., 2023). In recommendation, it can be the ability to steer retrieval toward watch-time objectives by conditioning the user tower on a watch-time variable $h$8 (Liu et al., 2024). In RAG, it may refer to balancing coverage, ranking quality, and factual consistency under entropy and variance constraints (Guo et al., 30 Oct 2025), or balancing answer accuracy against retrieval cost through an interpolation parameter $h$9 over retrieval strategies (Su et al., 17 Feb 2025).

This suggests that quality-controllable retrieval is better understood as a family of mechanisms rather than a single architecture. What unifies the family is the presence of an explicit control interface and an observable response in retrieval outputs or retrieval-conditioned generation.

2. Control targets and loci of intervention

The survey distinguishes several control dimensions relevant to retrieval. Under “What is controllable?”, it identifies Multi-Objective Control, Historical-Behavior Control, and Controllable Environmental Adaptation (Shen et al., 2024). Multi-Objective Control treats the target as a vector of objectives $f$0, such as precision, diversity, and latency, with a description vector $f$1 satisfying $f$2 (Shen et al., 2024). Historical-Behavior Control allows user-editable profiles or historical interaction manipulation to alter downstream quality. Environmental adaptation injects tags such as time-of-day or device context so that quality dimensions such as precision or freshness change with the environment (Shen et al., 2024).

The same survey also organizes control by where it is applied: pre-processing, in-processing, and post-processing (Shen et al., 2024). Pre-processing changes the input before retrieval. CtrlQGen is a canonical case: an entity is serialized as text, a control token indicating narrow or broad intent is prepended, and a T5-base model is trained to generate intent-conditioned synthetic queries that later reshape retriever training data or user query suggestions (Penha et al., 2023). The QCQC framework for text-to-image retrieval is another pre-processing design: a short underspecified query $f$3 is completed into $f$4 by a LLM conditioned on discretized relevance and aesthetic levels, after which a frozen VLM retrieves images using cosine similarity between $f$5 and image embeddings (Lu et al., 24 Feb 2026).

In-processing changes the retrieval model or routing dynamics themselves. Control-MVR learns separate self-supervised and supervised branches for audio and video, and mixes them at inference using $f$6 (Stewart et al., 2024). CRM inserts a scalar or vector condition $f$7 into the user tower of a two-tower retriever, making the scoring function conditional:

$f$8

(Liu et al., 2024). Flare-Aug interpolates the weights of two external classifiers,

$f$9

to decide whether a query should use no retrieval, single-step retrieval, or multi-step retrieval (Su et al., 17 Feb 2025). Multi-granular confidence-constrained RAG dynamically routes across multiple memory granularities and prunes low-confidence paths when entropy or variance thresholds are exceeded (Guo et al., 30 Oct 2025).

Post-processing alters ranking or retrieval outputs after a base model has produced candidates. The survey gives MMR as a classic example:

$f_{\mathcal{T}}$0

for trading relevance against redundancy (Shen et al., 2024). In code retrieval, CoQuIR describes multi-headed scoring in deployment: compute semantic score $f_{\mathcal{T}}$1 and quality score $f_{\mathcal{T}}$2, then combine them as $f_{\mathcal{T}}$3 (Geng et al., 31 May 2025). In generative text-music retrieval, GD-Retriever applies post-hoc control through negative prompting and DDIM inversion in the latent retrieval space rather than modifying the pretrained teacher encoder (Guinot et al., 22 Jun 2025).

3. Mechanisms of controllability

A major class of methods uses explicit interpolation between competing retrieval objectives. Control-MVR trains with self-supervised InfoNCE losses and supervised SupCon losses in both audio-to-video and video-to-audio directions, sums the four losses equally with $f_{\mathcal{T}}$4 in experiments, and then exposes a single inference-time knob $f_{\mathcal{T}}$5 to trade off broad audiovisual alignment against label information (Stewart et al., 2024). The same paper reports a smooth trade-off: at $f_{\mathcal{T}}$6 the model attains best self-supervised behavior with $f_{\mathcal{T}}$7 and $f_{\mathcal{T}}$8; at $f_{\mathcal{T}}$9 it reaches $\mathcal{X}$0 and $\mathcal{X}$1; intermediate values give $\mathcal{X}$2 at $\mathcal{X}$3 and $\mathcal{X}$4 at $\mathcal{X}$5 (Stewart et al., 2024).

A second class uses conditional generation to reshape the effective query. CtrlQGen learns $\mathcal{X}$6, where the condition $\mathcal{X}$7 specifies narrow or broad intent, and uses weak labeling to create triplets $\mathcal{X}$8 without manual intent annotations (Penha et al., 2023). The generated queries are then used either to retrain dense retrievers or to suggest broader user queries, reducing retrievability bias in Tracks, Podcasts, and Books (Penha et al., 2023). QCQC follows the same broad pattern but with different control semantics: it discretizes image relevance and aesthetics into levels such as Low, Medium, and High using percentile thresholds $\mathcal{X}$9 and $\mathcal{Y}$0, fine-tunes a LLM on captions conditioned by these levels, and then retrieves with a frozen OpenCLIP model (Lu et al., 24 Feb 2026).

A third class uses conditional features inside the retrieval model. CRM extends the two-tower paradigm by inserting a watch-time condition $\mathcal{Y}$1 into the user tower. In the DNN version, the conditional user representation is $\mathcal{Y}$2; in the transformer variant, cumulative watch-time-to-go tokens are fed into a Decision-Transformer-style block before producing $\mathcal{Y}$3 (Liu et al., 2024). The system uses a time-division multiplexing strategy at serving time, switching between $\mathcal{Y}$4 and $\mathcal{Y}$5 with probability $\mathcal{Y}$6, thereby steering the aggregate trade-off among watch time, video count, and interactions (Liu et al., 2024).

A fourth class uses routing, uncertainty, and closed-loop control. The multi-granular confidence-constrained RAG framework builds a hierarchical memory $\mathcal{Y}$7 across levels such as token, passage, document, and temporal or global memory, computes per-level routing weights

$\mathcal{Y}$8

fuses retrieved contexts, and regularizes generation with generation loss, entropy, and variance terms (Guo et al., 30 Oct 2025). During inference, if entropy or variance exceeds a threshold, the framework removes the lowest-confidence retrieval path and renormalizes the weights (Guo et al., 30 Oct 2025). MemR$\mathcal{Y}$9 implements a different closed-loop controller: a router selects among retrieve, reflect, and answer, while a global evidence-gap tracker maintains accumulated evidence $\Gamma$0 and unresolved gaps $\Gamma$1; deterministic constraints enforce an iteration budget and cap consecutive reflections (Du et al., 23 Dec 2025).

A fifth class uses generative latent-space manipulation. GD-Retriever generates a “ghost” audio embedding sequence in an audio-only latent space from text conditioning, retrieves nearest audio neighbors to the generated latent, and supports control through classifier-free guidance, negative prompting, and DDIM inversion (Guinot et al., 22 Jun 2025). Its negative prompting update explicitly repels an unwanted attribute embedding:

$\Gamma$2

(Guinot et al., 22 Jun 2025). This allows post-hoc manipulation of retrieval behavior without retraining the frozen audio teacher encoder.

4. Representative systems and application domains

The recent literature spans search, recommendation, cross-modal retrieval, code retrieval, and retrieval-augmented generation. The following systems illustrate the range of control variables and target quality notions.

System	Control variable	Controlled quality dimension
Control-MVR	$\Gamma$3	self-supervised vs. label-supervised retrieval
CtrlQGen	$\Gamma$4	retrievability and discovery
CRM	watch-time condition $\Gamma$5	watch time vs. video count vs. interactions
Flare-Aug	$\Gamma$6	accuracy vs. retrieval cost
QCQC	$\Gamma$7	relevance and aesthetics
CoQuIR-style retriever	$\Gamma$8 or dimension weights	correctness, efficiency, security, maintainability

In search, retrievability is treated as a measurable property of the catalog. CtrlQGen adopts the retrievability score

$\Gamma$9

with $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$0 in its experiments, and summarizes retrievability bias via the Gini coefficient $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$1 (Penha et al., 2023). On Tracks, a CtrlQGen-trained Bi-Encoder reduces Gini from $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$2 to $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$3 while increasing $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$4 from $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$5 to $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$6, and moves 9% of entities from zero to nonzero retrievability; broad-query suggestion lowers Gini by up to $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$7 and brings 11% more entities to $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$8 (Penha et al., 2023). These results situate quality control not only as re-ranking quality, but as control over the distribution of exposure across a catalog.

In recommendation, CRM treats retrieval quality as alignment between retrieval and downstream regression-oriented business targets. Offline, the transformer variant reaches 42 s average time per video view versus next best $$f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),$$9 s; online A/B tests on Kuaishou report significant gains including Video Watch Time: +0.372% / +0.457%, Total App Usage Time: +0.175% / +0.196%, and Follows: +0.447% / +0.828%, with all gains significant at $\mathbf{s}_{\mathrm{desc}}$0 (Liu et al., 2024). Here, controllability is embedded directly into the retrieval score rather than delegated to a later ranking stage.

In code retrieval, CoQuIR makes quality dimensions explicit and benchmarkable. It contains 42,725 natural-language–code queries, 134,907 candidate snippets, and 11 programming languages, and evaluates correctness, efficiency, security, and maintainability through Pairwise Preference Accuracy and Margin-based Ranking Score (Geng et al., 31 May 2025). Off-the-shelf models often achieve only $\mathbf{s}_{\mathrm{desc}}$1–$\mathbf{s}_{\mathrm{desc}}$2 and $\mathbf{s}_{\mathrm{desc}}$3–$\mathbf{s}_{\mathrm{desc}}$4 on many tasks, while quality-contrastive tuning improves PPA by 20–30 pp and MRS by 10–25 pp without sacrificing semantic relevance by more than about $\mathbf{s}_{\mathrm{desc}}$5–$\mathbf{s}_{\mathrm{desc}}$6 pp in nDCG@10/MRR (Geng et al., 31 May 2025). The benchmark thereby shifts quality control from an informal notion to an evaluable retrieval property.

In text-to-image retrieval, QCQC addresses short and underspecified queries by conditioning query completion on discrete relevance and aesthetic levels before retrieval with a frozen OpenCLIP ViT-H-14 model (Lu et al., 24 Feb 2026). On MS-COCO in the High-High condition, it improves relevance from 0.33 to 0.39 and aesthetics from 4.82 to 5.02 compared to the best fine-tuned unguided LLM, and five-level discretization still preserves smooth control (Lu et al., 24 Feb 2026). The paper also reports that off-the-shelf LLMs without fine-tuning often degrade retrieval quality and that post-retrieval reranking by aesthetics cannot jointly optimize relevance and aesthetics as effectively (Lu et al., 24 Feb 2026).

In retrieval-augmented generation and memory systems, quality control frequently targets reliability rather than only ranking quality. The multi-granular confidence-constrained RAG framework reports on CISI that its method attains 77.8 / 92.0 / 90.0 / 0.72 on QA Accuracy, Recall@5, NDCG@5, and factual consistency, compared with 69.3 / 80.0 / 78.0 / 0.65 for Self-RAG (Guo et al., 30 Oct 2025). MemR$\mathbf{s}_{\mathrm{desc}}$7, evaluated on LoCoMo, reaches 86.75% overall LLM-as-a-Judge score with a RAG backbone and GPT-4.1-mini, versus 79.46% for vanilla RAG, and improves both RAG and Zep backends while adding only 10–20% extra retrieval tokens relative to vanilla RAG (Du et al., 23 Dec 2025).

5. Evaluation methodology

Evaluation in quality-controllable retrieval is necessarily heterogeneous because the target quality dimension changes with the application. The survey lists standard single-objective IR metrics such as Precision@k, Recall@k, NDCG@k, Hit Rate, Coverage, $\mathbf{s}_{\mathrm{desc}}$8-NDCG, and ERR-IA, and multi-objective metrics such as Hypervolume and correlation between control parameter $\mathbf{s}_{\mathrm{desc}}$9 and an observed metric $\mathbf{s}_{\mathrm{ctx}}$0 via Pearson or Spearman correlation (Shen et al., 2024). It also prescribes an evaluation protocol in which one synthesizes a held-out requirement, generates the controlled model, measures target metrics, plots metric values against the control variable, and computes correlation or regret relative to ideal requirements (Shen et al., 2024).

Several systems add task-specific evaluation constructs. CtrlQGen uses retrievability score, Retrievability Gini, nonzero retrievability percentage, and $\mathbf{s}_{\mathrm{ctx}}$1 (Penha et al., 2023). Control-MVR distinguishes self-supervised retrieval, evaluated by Recall@K and MRR, from genre-supervised retrieval, evaluated by Precision@K and macro-averaged MRR over 11 genres (Stewart et al., 2024). CoQuIR introduces Pairwise Preference Accuracy, whose random baseline is 0.5 and ideal is 1.0, and Margin-based Ranking Score, where negative values indicate pathological preference for low-quality code (Geng et al., 31 May 2025). QCQC reports average relevance and average aesthetics over top-5 retrieved images (Lu et al., 24 Feb 2026). Flare-Aug uses answer accuracy and retrieval cost

$\mathbf{s}_{\mathrm{ctx}}$2

to characterize the accuracy–cost frontier (Su et al., 17 Feb 2025).

A recurring empirical criterion is monotonicity or smoothness of the response curve. Control-MVR reports a smooth trade-off over $\mathbf{s}_{\mathrm{ctx}}$3 (Stewart et al., 2024). Flare-Aug reports that both expected retrieval cost $\mathbf{s}_{\mathrm{ctx}}$4 and accuracy $\mathbf{s}_{\mathrm{ctx}}$5 are approximately monotonic in $\mathbf{s}_{\mathrm{ctx}}$6, and on Flan-T5 XL shows validation points such as $\mathbf{s}_{\mathrm{ctx}}$7 with accuracy 0.388 at 1.3 average steps, $\mathbf{s}_{\mathrm{ctx}}$8 with 0.439 at 2.2 steps, and $\mathbf{s}_{\mathrm{ctx}}$9 with 0.441 at 3.0 steps (Su et al., 17 Feb 2025). QCQC reports a clear monotonic trend across low-to-medium-to-high quality conditions (Lu et al., 24 Feb 2026). This emphasis on response curves is central: a controllable system is not merely high-performing, but predictably steerable.

6. Limitations, misconceptions, and open directions

A common misconception is that controllability is equivalent to a simple post-retrieval score adjustment. The surveyed systems contradict this. Control may be implemented by rule-based post-processing, Pareto optimization, hypernetworks, controllable query generation, conditional towers, uncertainty-based path pruning, or agentic retrieve–reflect–answer loops (Shen et al., 2024, Penha et al., 2023, Liu et al., 2024, Guo et al., 30 Oct 2025, Du et al., 23 Dec 2025). Another misconception is that natural-language prompting alone suffices. CoQuIR finds that pre-tuned instruction retrievers are largely insensitive, with <1 pp change in PPA/MRS, to explicit prompts such as “retrieve buggy” versus “retrieve correct,” indicating that specialized quality signals must be injected during training (Geng et al., 31 May 2025). QCQC likewise reports that off-the-shelf LLMs without fine-tuning often degrade retrieval quality (Lu et al., 24 Feb 2026).

The literature also shows that more retrieval is not automatically better. Flare-Aug is motivated by the fact that over-retrieving simple queries is wasteful and that users may require different points on the accuracy–cost spectrum (Su et al., 17 Feb 2025). MemR$h$00 explicitly prevents both under-retrieval and over-retrieval by maintaining evidence and gap sets and enforcing budgets such as $h$01 and $h$02 (Du et al., 23 Dec 2025). The confidence-constrained RAG framework states that in highly heterogeneous memories factuality degrades unless confidence constraints are tightened, which can in turn reduce coverage (Guo et al., 30 Oct 2025). These results indicate that controllability often mediates irreducible trade-offs rather than eliminating them.

Limitations recur across domains. The survey identifies balancing difficulty, absence of standardized evaluation, task description design, and real-time or online scalability as central challenges (Shen et al., 2024). Multi-granular RAG notes that computational overhead grows with the number of levels $h$03 and repeated pruning, and that $h$04, $h$05, and confidence thresholds require careful tuning (Guo et al., 30 Oct 2025). MemR$h$06 does not use explicit gradient-based training for its router or tracker, relying instead on prompt engineering and rule-based overrides (Du et al., 23 Dec 2025). GD-Retriever reports that diffusion sampling is significantly heavier than a single forward pass of a contrastive encoder and inherits distributional biases of its teacher encoders under domain shift (Guinot et al., 22 Jun 2025). QCQC reports that cross-dataset fine-tuning demonstrates the importance of dataset-specific adaptation for relevance control (Lu et al., 24 Feb 2026).

The research agenda identified by the survey includes theoretical analyses of controllable learners, cost-efficient control, LLM-empowered controllable learning, multi-task and multi-scenario switching, and shared benchmarks and metrics (Shen et al., 2024). The concrete systems surveyed here add more specific directions: extending discrete intent control to a continuous intent spectrum in CtrlQGen (Penha et al., 2023), enriching retrieval with multi-dimensional conditions in CRM (Liu et al., 2024), expanding multi-granular memory to cross-modal settings (Guo et al., 30 Oct 2025), and using active learning or multi-objective optimization for code-quality retrieval (Geng et al., 31 May 2025). Taken together, these directions frame quality-controllable retrieval as an increasingly general design principle for retrieval systems that must expose explicit, interpretable, and tunable behavior under changing user requirements and deployment constraints.