R1-SyntheticVL: Synthetic Data for MLLMs

• The paper introduces R1-SyntheticVL, a multimodal model that improves accuracy using synthetic adversarial data generated through the CADS framework.
• It employs a dual-phase process—CAD-Generate and CAD-Judge—with adversarial context optimization to iteratively refine data difficulty and quality.
• Empirical results on benchmarks like MathVista show notable gains, with up to a +7.4 percentage point improvement over the base model.

R1-SyntheticVL refers to a multimodal LLM trained with synthetic adversarial data synthesized using the Collective Adversarial Data Synthesis (CADS) framework. It is designed to enhance multimodal LLMs (MLLMs) on complex real-world benchmarks by leveraging entirely synthetic, challenging, and diverse multimodal data. The following sections systematically detail the technical aspects, methodologies, and empirical findings associated with R1-SyntheticVL as described in (Zhang et al., 3 Feb 2026).

1. Model Architecture

R1-SyntheticVL uses Qwen2.5-VL-7B as its backbone. The architecture consists of a vision encoder, which projects input images $v$ into continuous embeddings $\phi(v)$, and a standard transformer LLM that consumes both the visual embeddings and tokenized text input $q$. Visual embeddings are injected into the LLM via cross-modal attention layers, enabling unified vision-language modeling. The model outputs predicted answer tokens $\hat{a}$ given an image-question pair $(v, q)$. On top of the base Qwen2.5-VL-7B policy, a group-relative policy optimization (GRPO) loss is used in reinforcement learning to further optimize for higher correctness rewards on the synthesized training data.

2. Collective Adversarial Data Synthesis (CADS)

CADS formulates the data generation process as two cyclic phases, CAD-Generate and CAD-Judge, with an embedded Adversarial Context Optimization step.

2.1 CAD-Generate

Seeded with a set $\mathcal{D}_{\rm seed}$ of either extant multimodal problems or textual task prompts, a collective of $K$ LLMs $\{\pi_1, ..., \pi_K\}$ jointly generates diverse new multimodal instances:

• Rationale Analysis: Each $\pi_k$ extracts the problem’s domain and key reasoning steps.
• Synthesis-Strategy Generation: The collective chooses from four meta-strategies—parameter variation, logic reversion, auxiliary extension, and isomorphic-scenario transfer—to design new $(q',a')$ pairs.
• Visual-Prompt Generation: The abstracted problem is converted into an explicit text prompt encoding layouts, relations, and numeric values. This is rendered into an image $\phi(v)$0 using Nano Banana Pro.

2.2 CAD-Judge

Each generated instance $\phi(v)$1 is evaluated by all judges $\phi(v)$2. Each judge attempts to answer $\phi(v)$3 on $\phi(v)$4 to produce $\phi(v)$5. The agreement count $\phi(v)$6 determines retention:

• $\phi(v)$7: instance is discarded.
• $\phi(v)$8: marked “easy” or high confidence.
• $\phi(v)$9: considered “adversarial”—hard or ambiguous.

2.3 Adversarial Context Optimization

For partially agreed-upon samples ($q$0), disagreement patterns inform adaptive modification of synthesis strategy, refining prompts and strategy selection to maximize difficulty and value. The process iteratively increases the proportion of challenging, high-information data.

3. MMSynthetic-20K Dataset

Through CADS, the MMSynthetic-20K dataset is produced, comprising 20,000 high-quality, LLM-synthesized multimodal question-image-answer triples. Each includes a rendered image (via Nano Banana Pro), a textual question/instruction, and a short/free-form answer. Coverage is intentionally balanced across domains:

• Mathematics (geometry, algebra, arithmetic)
• Physics (mechanics, energy, collisions)
• Chemistry/Biology (molecular diagrams, biology schematics)
• Chart/Table comprehension (e.g., bar, line plots)

The teacher/judge ensemble comprises GPT-4o, Gemini-2.5-Flash, DeepSeek-R1, and Claude-4, iterated up to 10 times per seed for maximal diversity and robustness.

4. Training Framework and Loss Functions

R1-SyntheticVL is fine-tuned on MMSynthetic-20K using GRPO. Training details:

• Base policy: Qwen2.5-VL-7B
• Rollouts per prompt: 8
• Global batch size: 128 prompts
• Total rollout batch: 256 answers
• Learning rate: $q$1
• Hardware: 8× NVIDIA H20 GPUs, EasyR1 codebase

The loss function adopts a PPO-style (with group-normalized advantages as in GRPO):

$q$2

where $q$3 and $q$4 is the normalized advantage.

5. Empirical Evaluation and Benchmarking

R1-SyntheticVL is evaluated zero-shot against six established benchmarks, all reported as percentage accuracy:

Model	MathVista	MathVerse	MathVision	MMMU	MMMU-Pro Std-10	Vision	Reas.	Desc.	Avg
GPT-4o	63.8	50.2	30.4	70.7	54.0	49.7	47.1	84.5	56.3
R1-SyntheticVL	75.6	51.2	29.1	56.3	42.0	38.7	47.8	75.5	52.0

• On MathVista, R1-SyntheticVL achieves 75.6%, which is +7.4 percentage points over the same backbone without synthetic data.
• On MMMU-Pro, R1-SyntheticVL leads all open-source models on both standard and reasoning subtasks.
• The average across all tasks is 52.0%, the highest among large models trained with synthetic data (Zhang et al., 3 Feb 2026).

6. Ablation Studies

Ablations on MathVista detail the incremental accuracy impact of each CADS component:

Model variant	MathVista Accuracy
Qwen2.5-VL-7B (no synthetic data)	68.2
+ “Direct Nano Banana Pro” only	70.8
+ CAD-Generate	73.0
+ CAD-Generate + CAD-Judge	74.6
+ full CAD-Generate + CAD-Judge + Adv.	75.6

• CAD-Generate provides +2.2 points.
• CAD-Judge adds +1.6 points.
• Adversarial optimization contributes +1.0 points for the maximum observed gain.

These results underline the additive and essential role of all three CADS phases.

7. Context, Comparison and Limitations

R1-SyntheticVL establishes that synthetic data, when rigorously generated and filtered for difficulty and validity, can meaningfully boost MLLM performance, especially in symbolically rigorous domains such as mathematics and multimodal science. All claims, methodology, and metrics adhere strictly to those given in (Zhang et al., 3 Feb 2026). A plausible implication is that CADS-like processes may generalize to other domains requiring complex reasoning and robust generalization from synthetic supervision. However, the current design focuses on short-form answer generation and tasks with deterministic verification; thus, open-ended reasoning and generative outputs beyond answer scoring would require additional extensions.