Research map / Vision-language and multimodal / Multimodal LLMs

R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal Large Language Model?

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CADS, a multi-MLLM collective generate-judge loop with adversarial context optimization that synthesizes multimodal QA data (MMSynthetic-20K) via Nano Banana Pro, and trains R1-SyntheticVL with GRPO on Qwen2.5-VL-7B.

Problem

Synthesizing high-quality, diverse, and challenging multimodal training data for MLLMs is hard: directly using a image generator yields misaligned, repetitive, and trivial samples.

Approach

Use a group of MLLMs both to generate (rationale analysis, strategy generation, visual prompt generation) and to judge synthetic QA pairs via consensus; use inter-model disagreement instances as adversarial examples to refine the generation context over iterations.

Key results

• R1-SyntheticVL achieves 75.6 on MathVista versus 68.2 for the Qwen2.5-VL-7B base and 70.8 for directly applying Nano Banana Pro data.
• R1-SyntheticVL averages 52.0 across the six reported benchmarks, above ThinkLite-VL-7B (50.3) and other 7B reasoning baselines.
• Training on 2K synthetic samples gives 73.3 on MathVista versus 72.2 for 2K real MM-Eureka samples.

Surprising findings

• Directly applying Nano Banana Pro yields only marginal gains (68.2 to 70.8 on MathVista), while Stable Diffusion and Nano Banana reduce performance in the motivation table.

Limitations

• Relies on closed-source Nano Banana Pro for image generation, limiting reproducibility and cost transparency.
• Scaling experiment indicates no saturation at 20K, with the paper not exploring larger scales.
• Evaluation of synthetic data quality relies on consensus among MLLM judges without human verification reported in the main text.

Experimental setup

Datasets: MM-Eureka dataset MMSynthetic-20K

Benchmarks: MathVista MathVerse MathVision MMMU MMMU-Pro CharXiv

Metrics: accuracy

Baselines: GPT-4o Claude-3.5-Sonnet Kimi k1.5 LLaVA-OV-7B MiniCPM-V-2.6-8B LLaVA-OV-1.5-8B Mulberry-7B R1-VL-7B R1-Onevision-7B MMR1-7B-RL MM-Eureka-7B OpenVLThinker-7B

Builds on: GRPO Self-Instruct ScaleQuest EasyR1

Closest papers by meaning

• Learning from Reasoning Failures via Synthetic Data Generation (AAAI 2026 · similarity 0.86)
• It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs (NeurIPS 2026 · similarity 0.84)
• On Path to Multimodal Generalist: General-Level and General-Bench (ICML 2025 · similarity 0.83)
• Synthetic Text Generation for Training Large Language Models via Gradient Matching (ICML 2025 · similarity 0.83)
• CtrlSynth: Controllable Image Text Synthesis for Data-Efficient Multimodal Learning (ICML 2025 · similarity 0.83)
• MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans? (ICLR 2025 · similarity 0.83)