Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Jingyi Zhang, Tianyi Lin, Huanjin Yao, Xiang Lan, Shunyu Liu, Jiaxing Huang

TL;DR

R1-SyntheticVL addresses data scarcity in multimodal LLM training by proposing CADS, which combines collective generation by multiple MLLMs and collaborative judgment to synthesize high-quality, diverse, and challenging multimodal data. An Adversarial Context Optimization mechanism steers generation toward hard, informative samples, yielding MMSynthetic-20K. The synthetic data trains R1-SyntheticVL via GRPO, achieving state-of-the-art results on several benchmarks and showing potential to complement or reduce reliance on real data. The approach targets data bottlenecks in real-world tasks requiring precise visual-text alignment and multi-step reasoning, with practical implications for scalable MLLM development.

Abstract

In this work, we aim to develop effective data synthesis techniques that autonomously synthesize multimodal training data for enhancing MLLMs in solving complex real-world tasks. To this end, we propose Collective Adversarial Data Synthesis (CADS), a novel and general approach to synthesize high-quality, diverse and challenging multimodal data for MLLMs. The core idea of CADS is to leverage collective intelligence to ensure high-quality and diverse generation, while exploring adversarial learning to synthesize challenging samples for effectively driving model improvement. Specifically, CADS operates with two cyclic phases, i.e., Collective Adversarial Data Generation (CAD-Generate) and Collective Adversarial Data Judgment (CAD-Judge). CAD-Generate leverages collective knowledge to jointly generate new and diverse multimodal data, while CAD-Judge collaboratively assesses the quality of synthesized data. In addition, CADS introduces an Adversarial Context Optimization mechanism to optimize the generation context to encourage challenging and high-value data generation. With CADS, we construct MMSynthetic-20K and train our model R1-SyntheticVL, which demonstrates superior performance on various benchmarks.

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

TL;DR

R1-SyntheticVL addresses data scarcity in multimodal LLM training by proposing CADS, which combines collective generation by multiple MLLMs and collaborative judgment to synthesize high-quality, diverse, and challenging multimodal data. An Adversarial Context Optimization mechanism steers generation toward hard, informative samples, yielding MMSynthetic-20K. The synthetic data trains R1-SyntheticVL via GRPO, achieving state-of-the-art results on several benchmarks and showing potential to complement or reduce reliance on real data. The approach targets data bottlenecks in real-world tasks requiring precise visual-text alignment and multi-step reasoning, with practical implications for scalable MLLM development.

Abstract

In this work, we aim to develop effective data synthesis techniques that autonomously synthesize multimodal training data for enhancing MLLMs in solving complex real-world tasks. To this end, we propose Collective Adversarial Data Synthesis (CADS), a novel and general approach to synthesize high-quality, diverse and challenging multimodal data for MLLMs. The core idea of CADS is to leverage collective intelligence to ensure high-quality and diverse generation, while exploring adversarial learning to synthesize challenging samples for effectively driving model improvement. Specifically, CADS operates with two cyclic phases, i.e., Collective Adversarial Data Generation (CAD-Generate) and Collective Adversarial Data Judgment (CAD-Judge). CAD-Generate leverages collective knowledge to jointly generate new and diverse multimodal data, while CAD-Judge collaboratively assesses the quality of synthesized data. In addition, CADS introduces an Adversarial Context Optimization mechanism to optimize the generation context to encourage challenging and high-value data generation. With CADS, we construct MMSynthetic-20K and train our model R1-SyntheticVL, which demonstrates superior performance on various benchmarks.
Paper Structure (15 sections, 4 equations, 4 figures, 4 tables)

This paper contains 15 sections, 4 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multimodal Large Language Model
  4. Data Synthesis
  5. Method
  6. Analysis and Motivation
  7. Collective Adversarial Data Synthesis
  8. Collective Adversarial Data Generation
  9. Collective Adversarial Data Judgment
  10. Experiments
  11. Implementation Details
  12. Main Experimental Results
  13. Ablation Study
  14. Discussion
  15. Conclusion

Figures (4)

  • Figure 1: Comparison between "directly using Nano Banana Pro" and "our proposed Collective Adversarial Data Synthesis (CADS)". (a) Directly applying Nano Banana Pro often suffers from low data quality, limited data diversity and trivial difficulty in data synthesis of complex tasks, while (b) our proposed CADS effectively synthesizes high-quality, diverse and challenging multimodal data for MLLMs.
  • Figure 2: Overview of our proposed Collective Adversarial Data Synthesis (CADS). CADS operates with two cyclic phases, including (a) Collective Adversarial Data Generation (CAD-Generate), which leverages collective knowledge to generate diverse data and (b) Collective Adversarial Data Judgment (CAD-Judge), which collaboratively assesses the quality of the synthesized data. During these two phases, CADS introduces an (c) Adversarial Context Optimization mechanism to optimize the generation context and encourage challenging and high-value data generation.
  • Figure 3: Scaling analysis of synthetic data.
  • Figure 4: Qualitative illustrations of synthesized samples from MMSynthetic-20K.