DGTA: Dynamic Target Gen & Adaptation

• The paper demonstrates that DGTA frameworks enable dynamic target generation and multi-target adaptation through parameter-efficient modules, evidenced by improved FID, IS, and KID scores in image generation.
• DGTA is a framework that dynamically infers unknown targets or domains with minimal supervision, leveraging a shared backbone and lightweight adaptation modules in both vision and language tasks.
• Key methodologies include hyper-network based modulations for image synthesis and sequence-generative models for zero-shot stance detection, facilitating effective multi-target adaptation under few-shot conditions.

Dynamic Target Generation and Multi-Target Adaptation (DGTA) describes a class of frameworks and methodologies for adapting learned models to multiple, previously unseen targets or domains without requiring explicit enumeration of targets in advance. At its core, DGTA enables models to dynamically identify, represent, and adapt to diverse targets or domains, either in generation or understanding tasks, based only on limited supervision or data per target. The two principal instantiations—image generation via few-shot domain adaptation and natural language stance detection—demonstrate the generality of this approach, leveraging mechanisms such as hyper-networks, low-rank modulations, and sequence-generative LLMs for flexible and efficient adaptation (Kim et al., 2022, Li et al., 27 Jan 2026).

1. Formal Task Definitions and Core Principles

DGTA operates in settings where neither the number nor the identity of target domains (or stance targets) are known a priori. In image generation (DynaGAN), the objective is to modulate a pretrained generative model to produce samples in any of several domains $c \in \{e_1,\ldots,e_N\}$ given only a handful of examples per domain, while sharing as much structural knowledge as possible (Kim et al., 2022). In open-world stance detection, DGTA systems are required to produce the set

$\mathcal{Y}(x) = \{(t_i, s_i)\}_{i=1}^{N_x}$

for each input text $x$, where $N_x$ is unknown, each $t_i$ is a target entity, event, or concept span, and $$s_i \in \{\text{support},\text{against},\text{neutral}\}$$ is the stance (Li et al., 27 Jan 2026).

The underlying principles are:

• Dynamic target generation: targets are inferred or generated during inference rather than fixed at training.
• Multi-target adaptation: the system adapts to a variable (potentially large) set of targets/domains with minimal per-target parameter overhead, often relying on shared parameters or meta-parameterization.

2. Methodologies: Image Generation and Stance Detection

2.1. DynaGAN for Multi-Domain Few-Shot Image Adaptation

DynaGAN employs a fixed “backbone” StyleGAN2 generator $G$, augmented by an adaptation module $H$, a hyper-network parameterized by a target domain identifier $c$. This adaptation module produces trainable, layer-wise weight modulations—per-filter scale $\delta_l \in \mathbb{R}^{C_{\text{out}}}$ and rank-1-residual weights $\Delta\varphi_l$—allowing $G$ to synthesize diverse target domains while sharing most generator parameters across domains. The adaptation is realized by:

• Mapping $c$ via a small MLP $M_A$ to a code $\upsilon$.
• Predicting modulation parameters for each convolutional layer, with rank-1 tensor decomposition:

$$\Delta\varphi_l = \gamma_l \otimes \phi_l \otimes \psi_l$$

where only vectors $\gamma_l$, $\phi_l$, $\psi_l$ are predicted, yielding sublinear growth in adaptation parameters.

• At inference, changing $c$ switches the output domain; interpolating $c$ blends styles.

2.2. DGTA for Zero-Shot Stance Detection

For stance detection, two key sub-problems are solved:

• Target Generation: A fine-tuned sequence-generation LLM (e.g., Qwen2.5-7B) is prompted to extract all stance targets from a post $x$:

$$P(t_1,\ldots,t_{N_x} \mid x) = \prod_{j=1}^{L} P(w_j \mid w_{<j}, x)$$

with post-processing yielding the set of targets.

• Stance Assignment: For each identified $t_i$, another model (or the same, in end-to-end fine-tuning) predicts $s_i$ by maximizing $P(s_i \mid x, t_i;\theta_2)$.

Two model training regimes are explored:

• Integrated Fine-Tuning: Train a single LLM to output linearized $(t_i\colon s_i)$ pairs in sequence for a given text.
• Two-Stage Fine-Tuning: Decompose into separate target extraction and stance determination stages, each with dedicated objectives.

Both approaches employ LoRA adapter-based fine-tuning.

3. Loss Functions and Adaptation Techniques

3.1. Contrastive and Regularization Losses in DynaGAN

DynaGAN introduces a contrastive adaptation loss $\mathcal{L}_{\mathrm{contra}}$ to ensure that generated images from different domains are not only similar to the (few) real samples $\{I_c\}$ from their target, but also well-separated from those from other domains: $$\mathcal{L}_{\rm contra} = -\log \frac {\exp(\mathrm{sim}(E_{\mathrm{CLIP}}(\hat I_c(w)), E_{\mathrm{CLIP}}(I_c))/\tau)} {\exp(\mathrm{sim}(E_{\mathrm{CLIP}}(\hat I_c(w)), E_{\mathrm{CLIP}}(I_c))/\tau) + \sum_{j\ne c} \exp(\mathrm{sim}(E_{\mathrm{CLIP}}(\hat I_c(w)), E_{\mathrm{CLIP}}(\mathrm{Aug}(I_j)))/\tau)}$$ Additional losses include the Mind-the-Gap (CLIP+reconstruction) loss (from MTG) and an identity-preserving loss for face adaptation, yielding a joint objective: $$\mathcal{L} = \lambda_{\rm contra}\mathcal{L}_{\rm contra} + \lambda_{\rm MTG}\mathcal{L}_{\rm MTG} + \lambda_{\rm ID}\mathcal{L}_{\rm ID}$$ Only the adaptation module is trained; the backbone remains frozen.

3.2. Task Losses and Metrics in Stance Detection

DGTA models for stance tasks minimize maximum-likelihood losses for target set generation and stance assignment. Evaluation incorporates a comprehensive C-score for target accuracy: $$\text{C-Score} = (\alpha\,\mathrm{BERTScore} + \beta\,\mathrm{BLEU} + \gamma\,\mathrm{ROUGE\!-\!L}) \times \mathrm{Recall}$$ with fixed $\alpha, \beta, \gamma$. Stance label accuracy is computed via standard precision, recall, and $F_1$ metrics.

4. Evaluation Protocols and Empirical Results

4.1. Multi-Domain Few-Shot Generative Benchmarks

DynaGAN is evaluated on few-shot multi-target scenarios such as “cat$\to$animals” (10 species), “cat$\to$dogs” (5 images), or real$\to$artificial faces (9 styles). Quantitative results for cat$\to$animals report FID = 38.4 (vs. MTG-ext = 89.2), IS = 4.53 (vs. 2.37), and KID = 17.4 (vs. 58.5), demonstrating both improved fidelity and domain specificity compared to conventional baselines (Kim et al., 2022). DynaGAN achieves strong qualitative diversity and editability, avoiding domain averaging and preserving identity in the face domain.

4.2. Zero-Shot Stance Detection: Dataset and Metrics

A large-scale Chinese Weibo dataset is constructed: 70,931 posts (single-target: 27,148; dual: 25,312; triple: 9,835; multi-target $>$3: 8,636), with 72,705 support, 30,029 against, and 51,618 neutral stance labels (Li et al., 27 Jan 2026). Annotation combines three LLMs with cross-validation and human adjudication, yielding Fleiss’s $\kappa=0.82$ overall.

Key experimental outcomes:

• The two-stage Qwen2.5-7B achieves a C-Score of 66.99% for target identification.
• The integrated fine-tuned DeepSeek-R1-Distill-Qwen-7B attains stance detection $F_1$ of 79.26%.
• Fully fine-tuned models outperform prompted instruction models (e.g., GPT-4o, Llama3-8B) by 5–10 points.
• Performance degrades for cases with $N_x \ge 4$ targets (C-Score drops $<54\%$).

5. Computational Efficiency and Scalability

In few-shot image generation, standard per-domain fine-tuning requires $N\times G$ parameters for $N$ domains, whereas DynaGAN stores a single generator plus a lightweight adaptation module (e.g., 235.6M params for 10 domains vs. 303M for independent models), with negligible additional FLOPs due to rank-1 modulations. In stance detection, using LoRA adapters yields parameter-efficient fine-tuning, facilitating adaptation even as the number of discovered targets increases.

Scenario	Baseline Params	DynaGAN Params	Reduction
Cat$\to$Animals (10 domains)	303M	235.6M	$\sim$10x in aux models

This sublinear scaling is central to the efficiency advantage of DGTA models (Kim et al., 2022).

6. Limitations and Extensions

Documented limitations include reduced performance for “many-target” instances and difficulty in extracting or generating abstract/implicit targets (e.g., sarcasm, composite concepts) (Li et al., 27 Jan 2026). Semantic fragmentation during target generation and loss of inter-target context in two-stage stance detection are additional challenges.

Proposed extensions involve:

• Knowledge-augmented generation: retrieving external knowledge during target inference.
• Multi-stage reasoning: incorporating chain-of-thought or reinforcement learning for more robust stance and sense inference.
• Cross-lingual adaptation: aligning multilingual embeddings or constructing parallel datasets to generalize beyond Chinese.
• Graph-based models: explicitly modeling inter-target relations to jointly infer target sets and their interactions.

For image generation, future directions include generalizing hyper-network adaptation to modalities such as audio$\rightarrow$image or text$\rightarrow$image mapping, higher-rank decompositions for richer modulations, and continual/incremental learning to add domains without retraining the backbone.

7. Impact and Future Prospects

DGTA constitutes a foundational framework for open-world, zero-shot adaptation across multiple domains, both in generation (e.g., DynaGAN) and open-ended understanding tasks (e.g., stance detection). The key innovations—dynamic target discovery, shared/parameter-efficient adaptation modules, rank-1 modulations, and sequence-generation architectures—enable robust performance in challenging few-shot and multi-target regimes. Future work is likely to address abstract reasoning, richer relational modeling, and domain generalization, thereby extending DGTA’s applicability to increasingly open and complex real-world scenarios (Kim et al., 2022, Li et al., 27 Jan 2026).