CADS: Collective Adversarial Data Synthesis

• The framework formulates data synthesis as a bi-level min–max optimization, combining generator and judge cycles to iteratively improve data quality.
• It employs ensembles of multimodal LLMs to assess and filter synthetic data, ensuring diversity and adversarial difficulty in the resulting datasets.
• Empirical results on MMSynthetic-20K demonstrate that CADS boosts multimodal reasoning performance while reducing reliance on costly human annotations.

Collective Adversarial Data Synthesis (CADS) is a data synthesis framework developed for the autonomous construction of high-quality, diverse, and challenging multimodal datasets, specifically designed to advance training paradigms for Multimodal LLMs (MLLMs). CADS formulates data synthesis as a generator–judge loop, leveraging collective intelligence from ensembles of MLLMs to maximize the quality and adversarial difficulty of generated data. The core motivation is to produce synthetic data that drives substantial improvements in multimodal reasoning and generalization, mitigating the expense and limitations of human annotation at scale (Zhang et al., 3 Feb 2026).

1. Formal Objective and Optimization Structure

CADS frames synthetic data generation as a bi-level min–max optimization problem. The objective is to learn a generative policy $G_{\theta_G}$ producing a synthetic dataset $\mathcal{D}_{\mathrm{syn}}$ that exhibits high quality, diversity, and difficulty. A set of $K$ MLLMs, denoted $\Pi = \{\pi_1, \ldots, \pi_K\}$, acts as the collective "judge." Given a multimodal instance $(v', q', a')$ (image, question, answer), the consensus score is

$$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$

This partitions $\mathcal{D}_{\mathrm{syn}}$ into:

• Filtered dataset (solvable):

$$\mathcal{D}_{\mathrm{good}} = \{(v', q', a') \mid C(v', q', a') \geq 1\}$$

• Adversarial pool (challenging):

$$\mathcal{D}_{\mathrm{adv}} = \{(v', q', a') \mid 1 \leq C(v', q', a') < K\}$$

The generator is optimized by minimizing: $$\min_{\theta_G} \Big\{ L_{\mathrm{qual}}(\theta_G) + \alpha L_{\mathrm{div}}(\theta_G) + \beta L_{\mathrm{adv}}(\theta_G) \Big\}$$ where

• Quality loss:

$\mathcal{D}_{\mathrm{syn}}$0

• Diversity loss:

$\mathcal{D}_{\mathrm{syn}}$1

with $\mathcal{D}_{\mathrm{syn}}$2 as normalized (e.g. cosine) similarity.

• Adversarial difficulty loss:

$\mathcal{D}_{\mathrm{syn}}$3

where $\mathcal{D}_{\mathrm{syn}}$4 is a continuously optimized "generation context" vector.

This loss formulation ensures that only high-quality, broadly diverse, and adversarially difficult data is retained and iteratively improved.

2. Collective Adversarial Generation and Judgment Cycles

CADS operates through two cyclic phases—Collective Adversarial Data Generation (CAD-Generate) and Collective Adversarial Data Judgment (CAD-Judge):

CAD-Generate

A panel of strong MLLMs (e.g., GPT-4o, Gemini-2.5-Flash, DeepSeek-R1, Claude-4) co-generate candidate triples starting from a seed set $\mathcal{D}_{\mathrm{syn}}$5. For each seed $\mathcal{D}_{\mathrm{syn}}$6:

1. Rationale analysis: Each generator $\mathcal{D}_{\mathrm{syn}}$7 extracts its target knowledge domain $\mathcal{D}_{\mathrm{syn}}$8 (such as Geometry, Physics) and generates a chain of thought (CoT) $\mathcal{D}_{\mathrm{syn}}$9.
2. Synthesis-strategy construction: Based on $K$0, "meta-strategies" (e.g., Parameter-Variation, Logic-Reversion, Auxiliary-Extension, Isomorphic-Transfer) guide the creation of $K$1 candidates.
3. Visual-prompt generation: A textual prompt $K$2 is constructed to precisely specify a scene to the image generator (Nano Banana Pro).

The ensemble aggregates all generated candidates by majority vote or minimal similarity filtering, producing the batch $K$3. The process is formalized in the following pseudocode:

$\mathcal{D}_{\mathrm{syn}}$1

CAD-Judge

Each synthesized triple $K$4 is evaluated by the judge ensemble $K$5:

1. Every judge outputs $K$6.
2. The consensus score $K$7 is computed.
3. Only instances with $K$8 are retained ($K$9); cases where $\Pi = \{\pi_1, \ldots, \pi_K\}$0 are marked as adversarial ($\Pi = \{\pi_1, \ldots, \pi_K\}$1) for further context optimization. Instances with $\Pi = \{\pi_1, \ldots, \pi_K\}$2 are filtered out.

This two-phase loop iteratively sharpens data quality and adversarial content.

3. Adversarial Context Optimization

CADS introduces a continuously updated context vector $\Pi = \{\pi_1, \ldots, \pi_K\}$3 prepended to all visual prompts. The vector is optimized such that the image generator (Nano Banana Pro) produces scenes especially challenging for at least part of the judge ensemble. Optimization proceeds via:

$\Pi = \{\pi_1, \ldots, \pi_K\}$4

$\Pi = \{\pi_1, \ldots, \pi_K\}$5

Where $\Pi = \{\pi_1, \ldots, \pi_K\}$6 is the learning rate. The vector $\Pi = \{\pi_1, \ldots, \pi_K\}$7 is implemented as a low-rank soft prompt appended to the textual input to Nano Banana Pro. This mechanism increases the prevalence of "boundary-case" instances where the judge ensemble partially disagrees, thereby enhancing the overall adversarial value and informativeness of the dataset.

4. MMSynthetic-20K Dataset Construction

CADS was applied to synthesize MMSynthetic-20K, a high-entropy, 20,000-instance multimodal dataset. The construction process involved:

• Seed tasks: Drawn from MathVista, MMMU, CharXiv, and original textual prompts spanning geometry, physics, biology, and chart-based reasoning.
• Selection: Only entries with $\Pi = \{\pi_1, \ldots, \pi_K\}$8 post-judgment were retained. Near-duplicates (max-pairwise similarity $\Pi = \{\pi_1, \ldots, \pi_K\}$9 via CLIP embeddings) were pruned to maximize diversity.
• Category balance: Data distribution targeted $(v', q', a')$0 math, $(v', q', a')$1 physics, $(v', q', a')$2 biology, $(v', q', a')$3 charts. Empirical entropy of question-type label exceeded $(v', q', a')$4 bits over four meta-strategies.
• Preprocessing: All images were resized to $(v', q', a')$5, subjected to standard color-correction, and paired questions truncated to 128 tokens. Chain-of-thought (CoT) consistency was verified using a final LLM check.

Each instance in MMSynthetic-20K includes a $(v', q', a')$6 image, question text (with optional CoT), and the ground-truth answer.

5. Empirical Evaluation and Ablation

CADS demonstrates statistically significant performance gains on standard benchmarks. Key experimental findings include:

Setting	MathVista Accuracy (%)
Qwen2.5-VL-7B w/o synthetic data	68.2
+ direct Nano Banana Pro data	70.8
+ CAD-Generate only	73.0
+ CAD-Generate & CAD-Judge	74.6
+ Full CADS (+Adv. Context)	75.6

Further results:

• Closed-source/open-source comparison: R1-SyntheticVL trained solely on MMSynthetic-20K obtains $(v', q', a')$7 average on six vision-language benchmarks, outperforming all preceding open-source models. On MathVista, it scores $(v', q', a')$8 (compared to $(v', q', a')$9 for ThinkLite-VL-7B, $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$0 for Vision-R1-7B).
• Synthetic vs. real data efficiency: For MathVista, $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$1 real examples yield $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$2 accuracy, $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$3 synthetic (MMSynthetic) yield $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$4, and combining $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$5 (real+synthetic) produces $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$6.
• Scaling study: Performance on MathVista as a function of synthetic data size: $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$7, $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$8, $$C(v', q', a') = \sum_{k=1}^K \mathbf{1}[\pi_k(v', q') = a'].$$9, $\mathcal{D}_{\mathrm{syn}}$0.

These results support the utility of CADS for obtaining quality- and difficulty-calibrated synthetic data that rivals, or exceeds, the informativeness of expensive human-labeled datasets.

6. Broader Implications and Research Context

The collective adversarial formulation pioneered by CADS represents a convergence of multi-agent ensemble learning, synthetic data augmentation, and adversarial optimization in MLLM training regimes. Its generator–judge paradigm automates curriculum construction by targeting the "hard edge" of model agreement, systematically introducing complex and diverse problems into the training corpus. This suggests potential for application in domains beyond multimodal reasoning, wherever synthetic data generation and quality gating are needed.

A plausible implication is that CADS-like frameworks could standardize the production of high-fidelity, adversarial, and entropy-maximizing datasets as foundational MLLM resources, with direct impact on sample efficiency and task transferability in future models.

For detailed implementation and experimental protocols, consult "R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal LLM?" (Zhang et al., 3 Feb 2026).