Attribute Prompt Composition (APC)

• APC is a compositional framework that constructs prompts from reusable attribute-level components, enabling controlled prediction and generation.
• It leverages methods such as prompt algebra, fuzzy logic, and cross-attention to optimize modularity, generalization, and interpretability.
• APC improves performance in low-resource and out-of-distribution settings, as demonstrated in tasks like vision-language adaptation and entity matching.

Attribute Prompt Composition (APC) denotes a family of prompt design and optimization schemes in which prompts are built from reusable attribute-level components—such as semantic attributes, learned prompt vectors, factor texts, style tokens, or prompting techniques—and then combined to drive prediction, generation, retrieval, or reasoning. The term is explicit in object re-identification, where a Semantic Attribute Dictionary and Prompt Composition Module are used to form attribute-aware features (Wang et al., 23 Sep 2025), but closely related mechanisms also appear in low-resource entity matching, few/zero-shot vision-language adaptation, controllable generation, prompt optimization, and agent programming (Liu et al., 17 Mar 2026, Lee et al., 2024, Yang et al., 2022, Liu et al., 8 Apr 2026, Nath et al., 27 May 2026). Taken together, these works suggest that APC is best understood not as a single algorithm, but as a compositional design principle: represent attributes as modular prompt units, compose them with explicit operators, and optimize the resulting prompt program for controllability, generalization, or interpretability.

1. Conceptual scope and lineage

One line of work treats composability itself as the central object. À-la-carte Prompt Tuning trains one prompt per data source and composes them at inference time, so that models can be assembled from arbitrary selections of sources; the paper reports that such models achieve accuracy within $5\%$ of models trained on the union of the respective sources, with comparable training and inference cost, and state-of-the-art performance on Split CIFAR-100 and CORe50 (Bowman et al., 2023). Prompt Algebra formalizes another variant, asking whether independently learned prompts can be combined linearly to support composed tasks, with prompts constrained to a vocabulary-grounded subspace; its composite model is within $2.5\%$ of the best base model on average and improves UTZappos classification accuracy over the best base model by $8.45\%$ on average (Perera et al., 2023).

A second line makes attributes explicit. In low-resource entity matching, PROMPTATTRIB separates entity-level prompts from attribute-level prompts and composes attribute predictions by fuzzy logic (Liu et al., 17 Mar 2026). In few/zero-shot CLIP adaptation, CoAPT augments class prompts with LLM-generated attribute words, ATPrompt converts category-centric prompts into attribute-category hybrid prompts, and ArGue aligns prompts with primitive visual attributes while also introducing negative attributes to suppress spurious correlations (Lee et al., 2024, Li et al., 2024, Tian et al., 2023). In controlled generation, Tailor represents each attribute as a continuous prompt and composes multiple attribute prompts for multi-attribute text generation, while DCG learns attribute-oriented prompt vectors for unseen multi-attribute dialogue combinations (Yang et al., 2022, Zeng et al., 2023).

A third line generalizes APC beyond conventional prompt tuning. The object Re-ID APC framework defines a shared Semantic Attribute Dictionary and a Prompt Composition Module, with a Fast–Slow Training Strategy to balance discrimination and generalization (Wang et al., 23 Sep 2025). API-only prompt optimization frameworks such as aPSF factor prompts into semantic components and update one factor at a time (Liu et al., 8 Apr 2026), while Prompt Codebooks recast prompt optimization as routing each input to a small subset of discrete “instincts” from a codebook (Nath et al., 27 May 2026). This suggests that APC now spans both representation learning and prompt-program optimization.

2. Representational schemes for attributes

A common APC pattern is to embed attributes directly into the text sequence. ATPrompt replaces the standard class-only prompt

$P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$

with an attribute-category hybrid prompt such as

$$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$

where $[A]$ and $[B]$ are hard attribute tokens and the surrounding soft tokens are learned (Li et al., 2024). CoAPT uses the related form

$$\tilde{\mathbf{y}}^a_i = \{\mathbf{w}_{SOS}, \mathbf{p}^T_1, \cdots, \mathbf{p}^T_M, \mathbf{c}_i, \mathbf{a}_i^1, \cdots, \mathbf{a}_i^N, \mathbf{w}_{EOS}\},$$

so that soft prompt tokens, the class token, and LLM-generated attribute words are jointly encoded (Lee et al., 2024). ArGue likewise composes prompts from shared soft tokens, class tokens, and class-specific primitive visual attributes, then averages predictions across multiple attribute-conditioned prompts per class (Tian et al., 2023).

Another pattern represents attributes as learned prompt vectors rather than fixed words. Tailor learns one continuous prompt $S_k$ per controllable attribute and prepends it to a frozen GPT-2, while multi-attribute control is obtained by composing multiple single-attribute prompts (Yang et al., 2022). DCG constructs an attribute prompt text $P_{\text{att}}$ from attribute values, embeds it with the backbone LLM, and maps it through a shared MLP into attribute-oriented prompt vectors; these are then concatenated with task-oriented prompt tokens before the dialogue context (Zeng et al., 2023). Prompt Algebra makes this compositionality explicit with

$2.5\%$0

treating learned prompts as objects that can be combined through convex combinations (Perera et al., 2023).

A third pattern builds an explicit attribute inventory. In object Re-ID, APC defines a Semantic Attribute Dictionary

$2.5\%$1

with $2.5\%$2 learned attribute prompts by default and $2.5\%$3 learnable tokens per attribute in experiments; each attribute prompt is encoded by the frozen CLIP text encoder and later selected instance-wise (Wang et al., 23 Sep 2025). Prompt Codebooks use a finite codebook $2.5\%$4 of natural-language instincts, with practical $2.5\%$5, and assemble prompts from a routed subset of size $2.5\%$6 (Nath et al., 27 May 2026). aPSF similarly represents prompts as factor bundles $2.5\%$7 over a discovered factor schema $2.5\%$8, where factors correspond to roles such as task interpretation, reasoning procedure, verification, or output format (Liu et al., 8 Apr 2026).

3. Composition operators and control mechanisms

APC is not synonymous with naive concatenation. Tailor shows that simply concatenating single-attribute prompts can induce multi-attribute behavior, but also causes fluency decrease and position sensitivity; it therefore adds a Multi-Attribute Prompt mask to prevent cross-attention between prompts and a re-indexing position-ids sequence to eliminate order bias (Yang et al., 2022). This establishes one important APC principle: composition rules must control interference, not merely stack prompt tokens.

Other systems use explicit logical or algebraic operators. PROMPTATTRIB computes per-attribute probabilities for Same, Different, and Ambiguous, then aggregates them by fuzzy logic: $2.5\%$9

$8.45\%$0

with Ambiguous defined by max-plus-negation and final normalization over labels (Liu et al., 17 Mar 2026). Prompt Algebra instead composes prompts through convex linear combinations in a vocabulary-grounded subspace (Perera et al., 2023). In both cases, the composition operator is explicit and interpretable.

Attention-based routing is another APC mechanism. In object Re-ID, the Semantic Attribute Dictionary selects the Top-$8.45\%$1 attributes most similar to a projected visual representation $8.45\%$2, and the Prompt Composition Module performs two-stage cross-attention: first from selected attributes to image patches,

$8.45\%$3

then from the class token representation to attribute-conditioned features,

$8.45\%$4

yielding an attribute-aware feature $8.45\%$5 (Wang et al., 23 Sep 2025). AttriPrompt adopts a related but deeper mechanism: intermediate visual features are clustered into per-layer attribute vectors, used to retrieve semantically similar prompts from a prompt pool, and the retrieved prompts are concatenated to every layer of the text encoder (Zhan et al., 7 Sep 2025).

Prompt-program frameworks make composition procedural. aPSF assembles a prompt as

$8.45\%$6

then performs interventional single-factor updates by replacing one factor at a time and measuring the validation-score gain $8.45\%$7 (Liu et al., 8 Apr 2026). Prompt Codebooks route each input through an encoder to a subset of instincts, compose them with a generator, and update the encoder, generator, and codebook entries using critic-derived textual gradients (Nath et al., 27 May 2026). Adaptive Prompting for social bias detection treats prompt techniques themselves as attributes; its composition space is

$8.45\%$8

which equals $8.45\%$9 in the reported setup, and a DeBERTa-based selector predicts the best composition for each input (Spliethöver et al., 10 Feb 2025).

4. Representative realizations across domains

The breadth of APC is clearest when its instantiations are compared across tasks.

Area	Composed units	Composition mechanism
Low-resource entity matching	Attribute-level Same/Different/Ambiguous prompts	Fuzzy geometric mean, max, and negation
Few/zero-shot VLM adaptation	Class tokens, attribute words, soft prompts	Hybrid text prompts, sometimes with image-conditioned bias
Controlled generation	Single-attribute continuous prompts or attribute-oriented prompt vectors	Concatenation, masking, re-indexing, shared MLP composition
Object re-identification	Dictionary attributes and visual features	Top-$P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$0 retrieval plus cross-attention
API-only prompt optimization	Factors or instincts	Per-instance routing and factor-wise program assembly
Agent programming	Messages, tools, parsers, typed blocks	Declarative block composition in YAML

In entity matching, PROMPTATTRIB addresses a low-resource setting with only $P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$1 of labeled training data and shows that explicitly composing attribute signals is beneficial relative to entity-only prompting (Liu et al., 17 Mar 2026). In facial image editing, the 3D-aware Latent Attribute Editor composes prompts from learnable style tokens, system prompt words, and attribute text,

$P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$2

then maps the resulting text embedding to a latent editing direction for a frozen 3D-aware GAN (Kumar et al., 2024).

In vision-language classification, CoAPT fills unused text slots with attribute words and adds a meta-network that outputs image-specific bias vectors from concatenated image and text features (Lee et al., 2024). ArGue generates class-specific primitive visual attributes via GPT-3, samples the most visually relevant ones, and adds negative prompting with a shared class-agnostic attribute such as “the background of a” to flatten predictions on spurious features (Tian et al., 2023). ATPrompt searches over combinations of universal attributes such as color, shape, material, function, or habitat and embeds the selected attributes directly into the prompt as anchors (Li et al., 2024).

In generation, Tailor demonstrates that each controllable attribute can be represented by its own continuous prompt, and that multi-attribute generation can be improved by a trainable prompt connector together with position and attention corrections (Yang et al., 2022). DCG extends this to unseen multi-attribute dialogue combinations through shared attribute-oriented prompt vectors and a disentanglement loss (Zeng et al., 2023). In optimization, aPSF, PCO, and Adaptive Prompting all move APC toward instance-specific prompt programs: factors, instincts, or prompting techniques become selectable prompt attributes rather than fixed templates (Liu et al., 8 Apr 2026, Nath et al., 27 May 2026, Spliethöver et al., 10 Feb 2025).

5. Empirical behavior, generalization, and interpretability

Across domains, APC is primarily justified by improved out-of-distribution or low-resource behavior. PROMPTATTRIB improves over PromptEM on Geo-heter from F1 $P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$3 to $P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$4, on Cameras from $P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$5 to $P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$6, and improves accuracy on all four reported datasets, while a dropout-based contrastive variant raises Cameras F1 from $P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$7 to $P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$8 (Liu et al., 17 Mar 2026). In object Re-ID, APC achieves $P_T = [T_1][T_2]\dots[T_M][\mathrm{CLS}]$9 mAP / $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$0 Rank-1 on MSMT17, exceeds CLIP-ReID on DukeMTMC, Market1501, and VeRi-776, and on domain-generalized Market$$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$1MSMT17 reaches $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$2 versus $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$3 for CLIP-ReID (Wang et al., 23 Sep 2025).

In few/zero-shot CLIP adaptation, CoAPT raises the average base-to-novel harmonic mean from $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$4 to $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$5 for CoOp and from $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$6 to $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$7 for CoCoOp, while ATPrompt improves average harmonic mean over multiple textual prompt-learning baselines and shows consistent gains on base-to-novel, cross-dataset, and domain-generalization settings (Lee et al., 2024, Li et al., 2024). ArGue reports average harmonic mean $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$8 and ArGue-N $$P_T = [T_{a_1}]\dots[T_{a_m}][A]\;[T_{b_1}]\dots[T_{b_m}][B]\;[T_1]\dots[T_M][\mathrm{CLS}],$$9, exceeding LASP’s $[A]$0, and improves ImageNet-A from $[A]$1 with LASP to $[A]$2 and $[A]$3 with ArGue and ArGue-N respectively (Tian et al., 2023).

For controllable generation, Tailor achieves strong single-attribute control with only $[A]$4 training parameters of GPT-2 and reaches multi-attribute correctness $[A]$5 with the argmax-pseudo connector variant (Yang et al., 2022). DCG improves controllability on both DailyDialog-CG and ConvAI2-CG and, in seen-versus-unseen evaluations, exhibits much smaller degradation on unseen combinations than CTRL or naive prompt baselines (Zeng et al., 2023). Prompt-program APC shows similar trends. aPSF improves accuracy by up to $[A]$6 percentage points on average over strong prompt optimization baselines and reduces optimization cost by $[A]$7–$[A]$8 tokens on MultiArith while reaching peak validation in one step (Liu et al., 8 Apr 2026). Prompt Codebooks improve over zero-shot by up to $[A]$9 points, surpass GEPA by $[B]$0 on HotpotQA and $[B]$1 in aggregate, and reduce deployed prompt length by up to $[B]$2 versus MIPROv2 using only $[B]$3 instincts (Nath et al., 27 May 2026). Adaptive Prompting shows that prompt composition selection itself matters: on StereoSet with Llama-3, adaptive prompting reaches macro F1 $[B]$4 versus $[B]$5 for the best static composition (Spliethöver et al., 10 Feb 2025).

Interpretability is a recurring secondary benefit. PROMPTATTRIB exposes per-attribute probabilities and a transparent fuzzy aggregation path (Liu et al., 17 Mar 2026). Re-ID APC provides latent but inspectable attributes via dictionary entries, orthogonality, and instance-wise selection (Wang et al., 23 Sep 2025). aPSF yields explicit factor names and factor-level knock-out analyses (Liu et al., 8 Apr 2026), while PCO offers usage frequencies and success rates for individual instincts (Nath et al., 27 May 2026). This suggests that APC often improves performance precisely by making prompt structure more modular and auditable.

6. Limitations, misconceptions, and open directions

A persistent misconception is that APC is merely prompt concatenation. The literature argues otherwise. Tailor shows that naive concatenation introduces fluency decrease and position sensitivity unless masks and re-indexed positions are added (Yang et al., 2022). PROMPTATTRIB separates attribute prompts from entity prompts because multi-$[B]$6 formulations on long serialized inputs can exceed sequence limits, and then composes predictions with logic rather than token stacking (Liu et al., 17 Mar 2026). PCO and Adaptive Prompting further show that per-instance routing over a subset of components is structurally different from reusing one long monolithic prompt for all inputs (Nath et al., 27 May 2026, Spliethöver et al., 10 Feb 2025).

Several limitations recur. APC systems often depend heavily on attribute quality. PROMPTATTRIB assumes schema alignment and can degrade under noisy or misaligned schemas (Liu et al., 17 Mar 2026). ATPrompt relies on LLM-curated attribute bases and still uses fixed explicit attributes, while CoAPT’s gains weaken on some domain-generalization benchmarks because semantic attributes may not address style or texture shift (Li et al., 2024, Lee et al., 2024). In Re-ID, SAD effectiveness depends on the quality of the learned dictionary and orthogonality regularization, and attributes remain latent rather than text-labeled during training (Wang et al., 23 Sep 2025).

A second limitation is optimization overhead and search complexity. Adaptive Prompting requires evaluating a combinatorial composition space and training a selector per dataset and LLM (Spliethöver et al., 10 Feb 2025). aPSF depends on a capable Architect model and a small validation set, and its single-factor updates assume partial independence among factors (Liu et al., 8 Apr 2026). PCO introduces critic overhead and can suffer codebook collapse without explicit exploration (Nath et al., 27 May 2026). These results imply that APC trades off interpretability and modularity against a more elaborate optimization loop.

Open directions are already visible in the surveyed work. Multiple papers point toward learned or dynamic attribute discovery rather than fixed human- or LLM-specified attributes (Li et al., 2024, Zhan et al., 7 Sep 2025). Others suggest richer composition operators, including hierarchical or multi-head composition, learned logical templates, or sparsity-based attribute selection (Wang et al., 23 Sep 2025, Liu et al., 17 Mar 2026). The broader implication is that APC is evolving from a prompt-engineering heuristic into a general framework for modular prompt programs: attributes or factors are selected, composed, and revised as reusable units, with explicit credit assignment and instance-specific control.