Attribute-Conditioned Query Generator

• Attribute-conditioned query generators are models that synthesize queries controlled by specific semantic features and metadata, achieving high precision and interpretability.
• These systems utilize diverse conditioning techniques such as attribute tokens, cross-attention, latent variable concatenation, and prompt templating to fuse information effectively.
• They have been applied in various domains including e-commerce, security analysis, fine-grained image retrieval, and dialogue augmentation, demonstrating measurable improvements in retrieval and generation tasks.

An attribute-conditioned query generator is a model or algorithmic process that synthesizes queries (natural language, code, or structured representations) whose form and content are explicitly controlled or influenced by a set of conditioning attributes. These attributes may represent semantic features, task-specific labels, or metadata drawn from input contexts, and are introduced using explicit conditioning mechanisms within neural or symbolic architectures. The attribute-conditioned query generator paradigm underpins various recent advances in text generation, information retrieval, security analysis, and fine-grained representation learning, enabling high-precision, customizable, and interpretable query synthesis across diverse domains.

1. Fundamental Architectures and Conditioning Mechanisms

Attribute-conditioned query generation leverages a wide spectrum of architecture designs, each distinguished by how attributes are represented and fused with the query synthesis process.

• Neural generator–discriminator pipelines: In the mQE-CGAN framework (Cakir et al., 2022), query expansion employs a conditional generative adversarial network (CGAN) with a transformer-based sequence-to-sequence generator and a recurrent discriminator. Attribute conditioning vectors, derived from semantic insights (e.g., product metadata), are fused with the tokenized query either as prepended “attribute tokens” (input fusion) or as additional cross-attention keys/values (cross-attention bias). The discriminator concatenates a (projected) attribute vector to its hidden state, enforcing attribute-informed discrimination.
• Learnable query sets for attribute-aware hashing: For fine-grained vision retrieval (Wang et al., 21 Sep 2025), a set of k learnable queries is maintained, each corresponding to a specific attribute-aware hash bit. During decoding, transformer-style cross-attention forces each query to produce an attribute-specific feature, yielding binary codes that are both compact and interpretable.
• Conditional latent variable models: The CVAE architecture for dialogue query generation (d'Ascoli et al., 2019) conditions generation on discrete or structured attribute labels (e.g., intent class, slot vectors), concatenated alongside sampled latent codes at decoding. This approach enables diverse generation controlled by attribute vectors.
• Prompt engineering and tool-augmented agentic synthesis: In static analysis (security) query synthesis (Wang et al., 11 Nov 2025), attribute conditioning is implemented as structured prompt templates populated with CVE metadata fields, which direct both LLM synthesis and retrieval-augmented generation via external tool APIs and databases.

The conditioning fusion strategy—whether via prepended tokens, cross-attention, latent variable concatenation, or prompt templating—determines how attribute information is injected into the query generation process, influencing both fidelity and interpretability.

2. Representative Models and Task Domains

A variety of attribute-conditioned query generation models have been developed, each targeting domain-specific requirements:

Domain	Model/Framework	Attribute Types
Query Expansion (E-commerce)	mQE-CGAN (Cakir et al., 2022)	Product metadata (brand, category, material, price tier)
Static Analysis/Code Synthesis	QLCoder (Wang et al., 11 Nov 2025)	CVE fields (ID, description, CWE, severity, patch diff)
Fine-Grained Image Retrieval	QueryOpt (Wang et al., 21 Sep 2025)	Visual attributes (e.g., plumage, shape)
Dialogue Data Augmentation	CVAE (d'Ascoli et al., 2019)	Intent label, slot types

• In e-commerce, the mQE-CGAN model expands search queries conditioned on product attributes to increase retrieval efficacy by synthesizing attribute-relevant keywords, resulting in up to 9.8% higher semantic similarity compared to unconditioned baselines (Cakir et al., 2022).
• QLCoder targets code security analysis by synthesizing CodeQL queries from CVE metadata, where conditioning is achieved via structured prompt augmentation and retrieval-augmented generation, not via neural vector-embedding (Wang et al., 11 Nov 2025).
• QueryOpt learns a set of attribute-attending queries, each responsible for extracting attribute-specific information from dense visual embeddings, encoding it into interpretable and discriminative hash bits for retrieval (Wang et al., 21 Sep 2025).
• In low-resource dialogue, the CVAE-based generator produces intent-conditioned text queries, enabling controlled data augmentation without manual annotation (d'Ascoli et al., 2019).

3. Data Preparation, Attribute Representation, and Conditioning

Robust attribute-conditioned query generation depends on precise attribute extraction, normalization, and representation.

• Extraction and Encoding: Categorical attributes are generally one-hot encoded, while continuous attributes like prices are binned and embedded. Attribute vectors may be concatenated or summed to a fixed-dimension input (e.g., $c \in \mathbb{R}^d$), subsequently projected via multi-layer perceptrons or embedding layers (Cakir et al., 2022, Wang et al., 21 Sep 2025).
• Structural prompt templates: In code analysis, CVE fields are inserted verbatim into the code comment headers and tool-mediated prompts, aggregating all available condition-specific information to guide synthesis (Wang et al., 11 Nov 2025).
• Intent and slot representation: In the CVAE setting, attribute conditioning is over discrete class vectors, optionally augmented by slot presence, formality, or style features, supporting both categorical and multi-hot representations (d'Ascoli et al., 2019).
• Query set initialization and optimization: In vision models, each hash bit is aligned to a specific visual attribute by learning a set of queries $Q = [q_1; \dots; q_k]$, initialized randomly and optimized jointly with the network weights through cross-attention and pairwise similarity losses (Wang et al., 21 Sep 2025).

4. Training Objectives and Algorithms

Each framework introduces conditioning through its training loss or optimization loop:

• Conditional GAN (mQE-CGAN): Adopts a conditional minimax objective,

$$\min_{G}\,\max_{D}\; \mathbb{E}_{(x,c)\sim p_{\text{data}}}\big[\log D(x\mid c)\big] \;+\; \mathbb{E}_{z\sim p_z,\,c\sim p_{\text{data}}}\big[\log(1 - D(G(z\mid c)\mid c))\big]$$

with an auxiliary maximum-likelihood pretraining term for the generator (Cakir et al., 2022).

• Pairwise and quantization loss (QueryOpt): The loss combines a pairwise similarity term, aligning continuous outputs to the ground-truth similarity $S_{ij}$, and a quantization regularization penalizing the discrepancy between continuous and binarized hash codes (Wang et al., 21 Sep 2025).
• VAEs (CVAE): The loss is a sum of reconstruction, KL-divergence regularization, and categorical cross-entropy for attribute prediction, modulated by annealing and weighted for label/noise balancing in query transfer scenarios (d'Ascoli et al., 2019).
• Prompt-looped agentic feedback (QLCoder): No explicit backpropagation loss; instead, the agent synthesizes candidates, invokes external validation (compilation and run), and iteratively refines outputs based on tool feedback (Wang et al., 11 Nov 2025).

5. Evaluation Metrics and Empirical Results

Attribute-conditioned query generation is evaluated on several axes:

• Information retrieval and semantic similarity: Embedding-based cosine similarity, BLEU-4, and ROUGE-L are standard for expanded queries. mQE-CGAN reports up to 9.8% improvement in embedding similarity and 7–8% gains in BLEU/ROUGE metrics (Cakir et al., 2022).
• Hashing accuracy: QueryOpt is evaluated by mean average precision (mAP) across multiple datasets and bit-lengths, with attribute-aware queries yielding up to +15.8% mAP gains at low hash sizes compared to previous state-of-the-art (Wang et al., 21 Sep 2025).
• Intent-accuracy and diversity metrics: For dialogue, separate classifiers measure the match between generated query and conditioning intent; BLEU-diversity and originality (fraction novel vs. training set) assess diversity, with properly tuned transfer weights unlocking up to +20% in generation originality (d'Ascoli et al., 2019).
• Functional criteria: In static analysis, queries are assessed by compilation rate, recall, fix-discrimination, and F1 against vulnerable and patched code versions (Wang et al., 11 Nov 2025). QLCoder achieves 53.4% end-to-end success in CodeQL query synthesis, a substantial improvement over baseline LLM syntheses.

6. Empirical and Qualitative Examples

Concrete instantiations illustrate the conditioning process:

• E-commerce (mQE-CGAN): For a query “waterproof hiking boot” and attributes [category: footwear, subcategory: hiking boot, material: leather, price_tier: mid], the generator synthesizes “men’s leather waterproof hiking boots durable slip-resistant,” reflecting the injected attribute vectors (Cakir et al., 2022).
• Security analysis (QLCoder): For XXE CVE metadata, the system synthesizes a CodeQL query with annotation fields populated verbatim from the CVE (e.g., @security-severity 9.1), and synthesizes sanitization logic directly from patch diffs (Wang et al., 11 Nov 2025).
• Fine-grained image retrieval (QueryOpt): Analysis reveals that each learnable query attends to distinct visual attributes (e.g., beak or plumage in birds), and single-bit codes retrieve clusters of images sharing semantic traits (Wang et al., 21 Sep 2025).
• Dialogue generation (CVAE): Conditioning on “PlayMusic” intent, the CVAE model generates queries such as “Play [SongName] by [Artist],” with transfer-augmented sampling increasing novel utterance diversity without compromising intent adherence (d'Ascoli et al., 2019).

7. Extensions and Practical Considerations

Generalization to new tasks relies on modular attribute extraction, flexible fusion strategies, and task-suitable objectives.

• Multi-attribute and style conditioning: Both neural and prompt-based generators can condition on vectors representing multiple categorical or continuous attributes, supporting multi-intent, slot, style, or meta-attribute control (d'Ascoli et al., 2019).
• Auxiliary training branches and optimization: Auxiliary branches that transform conditioning queries (e.g., rotated query splits in QueryOpt) can alleviate optimization bottlenecks in low-dimensional settings (Wang et al., 21 Sep 2025).
• Adaptation to low-resource and transfer settings: Embedding-based retrieval (query transfer, RAG) can be combined with weak supervision, downweighted auxiliary data, or prompt engineering to inject attribute diversity (Wang et al., 11 Nov 2025, d'Ascoli et al., 2019).
• Evaluation and ablation: Proper benchmarking against unconditioned and attribute-agnostic baselines, along with visualization of attention patterns or attribute alignment, is crucial to validate the interpretability and discriminability of attribute-conditioned query generators across modalities and applications.

Attribute-conditioned query generators constitute a unifying paradigm for programmable, interpretable query synthesis, achieving substantial empirical gains in text expansion, information retrieval, code analysis, and controllable data augmentation by harnessing structured attribute signals at generation time (Cakir et al., 2022, Wang et al., 11 Nov 2025, Wang et al., 21 Sep 2025, d'Ascoli et al., 2019).