Target-Aware Representations in Feature Learning

• Target-aware representations are feature embeddings explicitly conditioned on designated target signals, guiding the learning process toward relevant structures.
• They employ mechanisms such as target-specific filter selection, cross-attention, and modality-aware fusion to tailor feature extraction across various domains.
• These methods enhance interpretability and performance by aligning features with task-specific targets and enabling efficient model adaptation.

Target-aware representations are feature embeddings explicitly conditioned on a designated target—such as an object instance, a linguistic referent, a support class, a downstream label, a protein target, or a transfer task—rather than generic task-agnostic signal. Across the literature, this conditioning is realized through target-specific filter selection, query initialization, cross-attention, target-token supervision, modality-aware fusion, contrastive alignment, or architecture adaptation, with the common objective of biasing representation learning toward the structures that matter for the target of inference (Li et al., 2019, Gu et al., 16 Feb 2025, Guo et al., 2022, Lan et al., 18 Sep 2025, Arazi et al., 23 May 2025).

1. Conceptual scope

In the surveyed works, “target-aware” does not denote a single architecture class. It denotes a family of mechanisms that inject explicit target information into representation learning. In visual tracking, the target may be the object specified in the first frame or the template branch of a Siamese network; in spatio-temporal grounding, it may be the subject and attributes extracted from a sentence; in drug discovery, it may be a protein target encoded with structural priors; in tabular learning and graph learning, it may be the supervised label space or downstream task itself (Li et al., 2019, Gu et al., 16 Feb 2025, Lan et al., 18 Sep 2025, Lin et al., 2024).

Setting	Target signal	Representative mechanism
Visual tracking	First-frame object or template	Gradient-based filter selection; target-aware cross-attention inside the backbone
Grounding and generation	Text subject, attributes, target mask, support/query relation	Target-aware query initialization; dependency-weighted heatmaps; target-token cross-attention
Drug discovery and molecular design	Protein target or interacting target sequence	Structure-aware tokenization; bilinear attention; latent alignment
Transfer, graphs, and tabular learning	Target dataset labels, class tokens, or downstream task	Activation-driven pruning; representation disentanglement; target-aware streams; task-aware contrastive sampling

A recurrent conceptual distinction is between generic representations, which are learned without explicit conditioning on the target, and target-aware representations, which alter the feature space itself. This distinction is explicit in visual tracking, where pre-trained CNN features are described as agnostic to unseen or arbitrary tracking targets, and in tabular modeling, where static target-agnostic textual representations are contrasted with semantically target-aware representations (Li et al., 2019, Arazi et al., 23 May 2025).

2. Instance-specific and template-conditioned tracking

In visual tracking, target-aware representations emerged as a response to the mismatch between generic recognition backbones and arbitrary tracked instances. “Target-Aware Deep Tracking” defines a target-aware representation as a low-dimensional, object-specific feature embedding obtained by selecting only those convolutional filters that are “active” in distinguishing the target from its background and sensitive to changes in the target’s scale (Li et al., 2019). The method uses a regression loss to fit a Gaussian “targetness” heatmap and a ranking loss over rescaled target patches. Per-filter importance is computed by globally pooling absolute back-propagated gradients, for example

$$\alpha_j^{reg}=\frac{1}{HW}\sum_{u,v}\left|\frac{\partial L_{reg}}{\partial A_{j,u,v}}\right|,$$

after which the tracker selects the top $K_{reg}=250$ filters from conv4-3 and the top $K_{rank}=80$ filters from conv4-1, concatenates them into $\chi'(\cdot)$, and matches template and search regions by cross-correlation. The resulting tracker reports OTB-2015 AUC $0.660$, OTB-2013 AUC $0.680$, VOT-2015 EAO $0.327$, and $33.7$ FPS on a GTX-1080 GPU (Li et al., 2019).

A later line of work moves target awareness from post hoc feature selection to backbone-internal interaction. “Learning Target-aware Representation for Visual Tracking via Informative Interactions” argues that in standard Siamese trackers only the matching module directly accesses target information, leaving shallow candidate-frame features blind to the reference target (Guo et al., 2022). Its Interaction-inside-Backbone mechanism inserts a General Interaction Modeler after multiple stages, with Window Process, Context-aware Self-Attention, and Target-aware Cross-Attention. In the cross-attention step, template information $z$ is injected into the candidate stream $x$, which boosts target-matching candidate features and suppresses distractors. On LaSOT and TNL2K, the CNN version improves the SiamCAR baseline from SUC $K_{reg}=250$0 to $K_{reg}=250$1 and from $K_{reg}=250$2 to $K_{reg}=250$3, while the Transformer version reaches SUC $K_{reg}=250$4 on LaSOT/TNL2K; the reported overhead is $K_{reg}=250$5 on ResNet-50 and less than $K_{reg}=250$6 on the Transformer, preserving real-time inference (Guo et al., 2022).

In RGBD tracking, target-aware representation learning is formulated as modality-aware fusion rather than single-stream specialization. DMTracker first learns modality-shared features through a Cross-Modal Integration Module and then reinjects RGB-specific and depth-specific information with a Specificity Preserving Module (Gao et al., 2022). The shared feature $K_{reg}=250$7 is derived by cross-attention between RGB and depth tokens, and the final representation is formed as

$K_{reg}=250$8

with $K_{reg}=250$9. On DepthTrack the tracker reports $K_{rank}=80$0, $K_{rank}=80$1, $K_{rank}=80$2, and on CDTB it reports $K_{rank}=80$3, $K_{rank}=80$4, $K_{rank}=80$5, surpassing the cited DeT baseline (Gao et al., 2022).

Taken together, these works show two non-equivalent meanings of target awareness in tracking: explicit instance-specific feature selection in the first frame, and repeated template-conditioned feature shaping throughout the backbone.

3. Query-aware grounding, few-shot localization, and target-aware generation

In spatio-temporal video grounding, target-aware representations are used to replace zero-initialized object queries. TA-STVG introduces Text-Guided Temporal Sampling and Attribute-Aware Spatial Activation as a cascade that generates object queries directly from the given video-text pair (Gu et al., 16 Feb 2025). TTS computes frame relevance scores by fusing appearance and motion scores with $K_{rank}=80$6, thresholds the resulting $K_{rank}=80$7, and passes target-relevant temporal cues to ASA. ASA then extracts a subject embedding by dependency parse, uses cross-attention over sampled appearance and motion features, supervises attribute prediction with $K_{rank}=80$8, and uses the attention weights as spatial activation maps. The initial object queries $K_{rank}=80$9 and

$\chi'(\cdot)$0

thus already carry target-specific cues before entering the decoder (Gu et al., 16 Feb 2025).

Few-shot learning adopts a related but distinct strategy. SpatialFormer’s SpatialFormer Target Attention sets the reference object $\chi'(\cdot)$1 to the global-classifier weight matrix $\chi'(\cdot)$2, so that support and query feature maps attend to base-class weight vectors as a proxy for “what objects look like” in the global feature space (Lai et al., 2023). The module computes $\chi'(\cdot)$3, then reweights each spatial query by a cosine-similarity-derived factor, amplifying patches that align with base-class object prototypes while leaving low-similarity background patches relatively unenhanced. On miniImageNet 5-way 1-shot, the ablation table reports $\chi'(\cdot)$4 for ProtoG, $\chi'(\cdot)$5 for SFTA only, and $\chi'(\cdot)$6 for the combined STA block (Lai et al., 2023).

Weakly supervised visual grounding uses target awareness at both training and inference. “Focusing On Targets For Improving Weakly Supervised Visual Grounding” introduces target-aware cropping, in which a proposal box $\chi'(\cdot)$7 obtained from a warm-up model is interpolated with the full-image box $\chi'(\cdot)$8 via $\chi'(\cdot)$9, with $0.660$0 and best $0.660$1 (Pham et al., 2023). At inference, it replaces uniform averaging of token Grad-CAM maps with a dependency-weighted sum in which tokens on or before the root receive weight $0.660$2 and later tokens receive weight $0.660$3, with best $0.660$4. The reported gains include RefCOCO val $0.660$5, RefCOCO testA $0.660$6, RefCOCO+ val $0.660$7, and RefCOCOg val(G) $0.660$8 for X-VLM plus the proposed modules (Pham et al., 2023).

Target-aware generation extends the concept from representation learning for recognition to representation learning for control. TAVDM adds a binary mask channel to CogVideoX-5B-I2V and appends the sentence “The person interacts with [TGT] object.” to the prompt, learning a new $0.660$9 token while freezing the rest of the pretrained weights apart from LoRA adapters and the new mask-projection layer (Kim et al., 24 Mar 2025). A cross-attention loss aligns the video-to-text attention map associated with $0.680$0 to the downsampled target mask:

$0.680$1

with $0.680$2 and supervision restricted to $0.680$3. On a 50-image benchmark with 5 samples each, the Contact Score rises from $0.680$4 for CogVideoX to $0.680$5 for the full method (Kim et al., 24 Mar 2025).

These systems share a common principle: the query, referent, or control target is not merely decoded from a generic feature map after the fact; it is injected into the representation before or during feature formation.

4. Biological target conditioning in drug discovery and molecular design

In computational pharmacology, target-aware representations are designed to preserve the scalability of sequence-based methods while encoding binding-relevant structure. SaBAN-DTI constructs target-aware protein embeddings by injecting structural priors into plain sequences, learning to focus on binding-relevant regions, and aligning those representations with drug embeddings via contrastive learning (Lan et al., 18 Sep 2025). Protein sequences are tokenized with the SaProt vocabulary of $0.680$6 tokens that jointly encode residue identity and local 3D geometry; drug molecules are represented as SELFIES; frozen Saprot and SELFormer encoders produce token-level embeddings; attention pooling yields global summaries; and a bilinear attention network captures fine-grained ligand–residue contacts. In parallel, pooled drug and protein embeddings are projected to a shared $0.680$7-dimensional space and optimized with a symmetric InfoNCE loss

$0.680$8

The reported results are state-of-the-art AUROC on Human $0.680$9 and BioSNAP $0.327$0, competitive BindingDB AUROC $0.327$1, and LIT-PCBA AUROC $0.327$2 with BEDROC $0.327$3. The ablations identify learned aggregation as the largest contributor to AUROC/AUPRC, BAN as critical for BEDROC and EF, and contrastive learning as important for generalization to unseen targets; t-SNE and attention visualizations show concentration on known binding-pocket token patterns and residues annotated as pocket-lining or catalytically important (Lan et al., 18 Sep 2025).

Target-aware molecular generation addresses a different problem: generating molecules conditioned on a target sequence. SiamFlow aligns a flow-based molecular graph latent $0.327$4 with a protein-sequence embedding $0.327$5 instead of assuming a standard Gaussian prior (Tan et al., 2022). The core objective combines an alignment loss, $0.327$6, with a hyperspherical uniformity loss over normalized target embeddings, and then extends alignment to a one-to-many setting by sampling from a Gaussian ball $0.327$7. The benchmark contains $0.327$8 unique $0.327$9 pairs, $33.7$0 distinct molecules, and $33.7$1 proteins with zero protein overlap among train, validation, and test. Reported generative metrics are $33.7$2 validity, $33.7$3 uniqueness, and $33.7$4 novelty for SiamFlow, compared with $33.7$5 for Seq2seq and $33.7$6 for CVAE; without one-to-many alignment, uniqueness collapses to $33.7$7 (Tan et al., 2022).

In this domain, target awareness means conditioning molecular or protein representations on the biological interaction partner rather than on chemistry alone.

5. Target-aware transfer learning and architecture adaptation

A separate line of work uses target awareness to reshape models for a new target task or target domain. “Target Aware Network Adaptation for Efficient Representation Learning” defines target-aware transfer learning as automatic adaptation of a pre-trained ConvNet architecture to the target dataset by iteratively pruning under-activated filters and re-optimizing (Zhong et al., 2018). For each layer, it computes channel-wise activation averages, normalizes them, forms a cumulative sum vector $33.7$8, chooses the smallest $33.7$9 approximating a threshold ratio $z$0, and scores the layer by

$z$1

Layers with $z$2 below the mean priority are pruned. On VGG-16, the method improves fine-tuning from $z$3 to $z$4 on CUB200, from $z$5 to $z$6 on MIT67, and from $z$7 to $z$8 on Stanford40 while reducing parameters and FLOPs substantially; for CUB200, the table reports parameters $z$9 and FLOPs $x$0 relative to the original $x$1 (Zhong et al., 2018).

TRED reframes target awareness as disentanglement of transferable and non-transferable source knowledge (Li et al., 2020). A frozen source network produces $x$2, a lightweight disentangler splits it into $x$3 and $x$4, and the decomposition is trained to satisfy distinguishability, discriminativeness, and recoverability. Two variants are provided: Max-MMD, which maximizes discrepancy between spatial summaries, and Min-MI, which minimizes mutual information between channel summaries using MINE. Fine-tuning then regularizes the target model against the positive feature map only: $x$5 Across seven datasets, the table reports average top-1 accuracy $x$6 for TRED-MMD and $x$7 for TRED-MI, compared with $x$8 for DELTA and $x$9 for $K_{reg}=250$00-SP; the abstract states that the method stably improves standard fine-tuning by more than $K_{reg}=250$01 in average (Li et al., 2020).

In whole-slide image classification, TAKT uses a teacher–student framework in which the teacher learns from the source domain while unlabeled target images provide target features for adaptation (Xiong et al., 2023). Its Target-Aware Feature Alignment module applies Power–Temperature Scaling with $K_{reg}=250$02 and $K_{reg}=250$03, an $K_{reg}=250$04 multi-head projection, and gated attention pooling to map teacher bag features into the student feature space. The paper also presents an optimal transport view with entropic regularization $K_{reg}=250$05. In transfer settings, the reported AUCs are $K_{reg}=250$06 for TCGA-RCC $K_{reg}=250$07 TCGA-NSCLC, $K_{reg}=250$08 for TCGA-NSCLC $K_{reg}=250$09 TCGA-RCC, $K_{reg}=250$10 for TCGA-RCC $K_{reg}=250$11 Camelyon16, and $K_{reg}=250$12 for TCGA-NSCLC $K_{reg}=250$13 Camelyon16, each exceeding the listed baseline and fine-tuning alternatives (Xiong et al., 2023).

These works broaden the term beyond object- or query-conditioned inference: the “target” can be the target dataset itself, and target-aware representations can be produced by pruning, disentanglement, or feature-space alignment.

6. Label-aware and task-aware representations in graphs and tabular data

In tabular supervised learning, target-aware representations can be explicitly label-conditioned. SG-XDEAT maintains two parallel views of each feature: a raw value stream $K_{reg}=250$14 and a target-aware stream $K_{reg}=250$15, where categorical features are transformed via shallow decision trees trained to predict $K_{reg}=250$16, and numerical features are partitioned by PLE-T (Cheng et al., 14 Oct 2025). Cross-encoding self-attention fuses global encoding tokens, raw tokens, and target-aware tokens on a per-feature basis, while Adaptive Sparse Self-Attention mixes a softmax branch with a squared-ReLU branch through learned coefficients $K_{reg}=250$17. The reported input-strategy ablation on California Housing gives RMSE $K_{reg}=250$18 for raw alone, $K_{reg}=250$19 for target-aware alone, $K_{reg}=250$20 for direct concatenation, and $K_{reg}=250$21 for the full SG-XDEAT model; the main-results table reports Adult accuracy $K_{reg}=250$22 versus $K_{reg}=250$23 for FT-Transformer and California Housing RMSE $K_{reg}=250$24 versus XGBoost’s $K_{reg}=250$25 (Cheng et al., 14 Oct 2025).

Target awareness in graph representation learning is formulated in terms of mutual information with the downstream task. XTCL defines a Task-Aware Contrastive Loss in which positive samples for a node $K_{reg}=250$26 are chosen by an XGBoost-style sampler that estimates

$K_{reg}=250$27

from multiple graph relations (Lin et al., 2024). Under the InfoNCE argument described in the paper, if the positives approach the true task-positive distribution, minimizing XTCL increases the mutual information between the target task and node embeddings. The reported node-classification accuracies include XTCL(GCN) $K_{reg}=250$28 on Cora, $K_{reg}=250$29 on CiteSeer, and $K_{reg}=250$30 on Computers, while link-prediction AUC includes $K_{reg}=250$31 on Cora, $K_{reg}=250$32 on CiteSeer, and $K_{reg}=250$33 on Photo (Lin et al., 2024).

TabSTAR extends target-aware representation learning to foundation-style tabular models with text features (Arazi et al., 23 May 2025). Its defining move is to treat every possible target value as an input token: for classification, target verbalizations such as “Target Dec : $K_{reg}=250$34” are concatenated with feature verbalizations, encoded by an unfrozen e5-small-v2 encoder, fused with numerical projections, and contextualized by a Transformer without positional encodings. The paper defines a semantically target-aware representation as a contextual embedding that jointly encodes feature values and every possible target value. Because the classification head is shared across class tokens and datasets, the architecture is described as having zero dataset-specific parameters. On the 10K classification benchmark, TabSTAR reports $K_{reg}=250$35, versus $K_{reg}=250$36 for TabPFN-v2, $K_{reg}=250$37 for CatBoost-Tuned, and $K_{reg}=250$38 for XGBoost-Tuned; under the unlimited condition, TabSTAR-Unlim reports $K_{reg}=250$39 (Arazi et al., 23 May 2025).

These methods make explicit that target-aware representation learning need not be unsupervised or weakly supervised. In some settings, the label space itself is an integral part of the representation.

7. Recurrent design patterns, interpretability, and limitations

Several design patterns recur across otherwise disparate application areas. One pattern is explicit target injection at the input level: SaBAN-DTI uses a structure-aware vocabulary; TAVDM adds a mask channel and a $K_{reg}=250$40 token; TabSTAR prepends target verbalizations; SG-XDEAT constructs a target-aware token stream (Lan et al., 18 Sep 2025, Kim et al., 24 Mar 2025, Arazi et al., 23 May 2025, Cheng et al., 14 Oct 2025). A second pattern is target-conditioned interaction: InBN injects template information via Target-aware Cross-Attention, TA-STVG generates target-aware queries through TTS and ASA, SpatialFormer attends to base-class weights, and DMTracker fuses modalities through cross-modal attention (Guo et al., 2022, Gu et al., 16 Feb 2025, Lai et al., 2023, Gao et al., 2022). A third pattern is alignment or selection: XTCL changes which positives are selected, TRED isolates the positive part of source representations, SiamFlow aligns molecular and target-sequence latents, and SaBAN-DTI aligns pooled drug and protein embeddings with a symmetric InfoNCE loss (Lin et al., 2024, Li et al., 2020, Tan et al., 2022, Lan et al., 18 Sep 2025).

Interpretability is also a recurrent consequence of target-aware formulations. In Target-Aware Deep Tracking, the absolute back-propagated gradients rank target-active and scale-sensitive filters; in weakly supervised grounding, dependency-based weights shift Grad-CAM emphasis from distractor words to referent words; in SaBAN-DTI, pooling weights and BAN attention concentrate on pocket-lining or catalytically important residues; in XTCL, XGSampler weights reveal which relations are predictive for node classification versus link prediction (Li et al., 2019, Pham et al., 2023, Lan et al., 18 Sep 2025, Lin et al., 2024).

A common misconception is that target awareness necessarily implies heavy online adaptation or dense control signals. The surveyed literature shows otherwise. Target-Aware Deep Tracking performs no further online learning at test time; InBN reports only small overhead; SaBAN-DTI keeps frozen encoders and feeds only token indices at inference time; and TAVDM requires only a simple mask rather than dense structural or motion cues (Li et al., 2019, Guo et al., 2022, Lan et al., 18 Sep 2025, Kim et al., 24 Mar 2025). Another misconception is that “target-aware” always refers to object-centric perception. The transfer-learning, tabular, and graph papers show that the target may be a downstream task, a label-aligned stream, or a target-domain feature manifold rather than a localized object (Zhong et al., 2018, Li et al., 2020, Xiong et al., 2023, Arazi et al., 23 May 2025).

The limitations are correspondingly heterogeneous. SpatialFormer notes that if novel classes are truly out-of-distribution relative to base classes, $K_{reg}=250$41 may not highlight their foreground accurately; XTCL requires a predefined set of semantic relations and can incur $K_{reg}=250$42 cost if all pairwise relations are enumerated; TAVDM identifies extensions to multiple masks and complex scenes with small or cluttered targets; TabSTAR notes that memory scales with the number of feature tokens in very wide tables (Lai et al., 2023, Lin et al., 2024, Kim et al., 24 Mar 2025, Arazi et al., 23 May 2025). This suggests that “target-aware representation” is best understood not as a solved recipe but as a design principle: the representation is deliberately reparameterized around the target signal, and the practical trade-offs depend on which target signal is available and how directly it can supervise feature formation.