Yuvion VL: A Multimodal Foundation Model for Adversarial Content and AI Safety

2026-06-23

AI Summary

This work addresses the challenge that existing general-purpose multimodal models struggle to reliably detect adversarial risk content in real-world scenarios. Framing AI safety as an inherently adversarial and multimodal problem, the authors propose Yuvion VL-32B, an end-to-end robust multimodal large language model with two variants—instruction-tuned and reasoning-enhanced. They introduce a novel "Confuse-then-Contrast" fine-tuning strategy that first identifies model-confusing samples and then constructs multi-image contrastive groups to enhance discrimination of fine-grained visual-semantic differences. For the first time, adversarial robustness is integrated throughout the entire pipeline of a multimodal safety model. Through a three-stage training regimen combining adversarial-aware data synthesis, cross-modal alignment, and contrastive learning, Yuvion VL-32B achieves superior safety performance over both open-source and leading closed-source models of comparable scale while preserving general capabilities, and introduces Yuvion VL RiskEval, a multidimensional safety evaluation benchmark.

Abstract

General-purpose models often struggle to reliably identify and understand real-world multimodal risks, largely due to the inherent multimodal adversarial nature of content and AI safety. We present Yuvion VL, a family of multimodal large language models purpose-built for content and AI safety, with both instruction-tuned and reasoning-oriented variants. Yuvion VL addresses this gap by treating safety as an inherently adversarial and multimodal problem and designing the entire pipeline around adversarial robustness. For data construction, we develop an automated pipeline integrating adversarial-aware data synthesis with multi-stage quality control, producing large-scale, high-quality multimodal samples augmented with domain knowledge and reasoning annotations. For training, we adopt a three-stage pipeline that includes continued pretraining for risk-concept cross-modal alignment, instruct post-training for production-grade safety tasks, and reasoning post-training for enhanced interpretability and performance in complex tasks. We further introduce Confuse-then-Contrast Fine-Tuning, a contrastive framework that mines model-specific confusions and constructs multi-image contrastive groups to enforce explicit discrimination of fine-grained visual-semantic elements, enabling the model to distinguish between visually similar cases with different safety implications in adversarial safety tasks. To support rigorous evaluation, we further introduce Yuvion VL RiskEval (YVRE), a collection of benchmarks covering diverse open and internal evaluations, with a focus on content and AI safety, adversarial robustness, and real-world capability requirements. Experiments show that Yuvion VL-32B achieves industry-leading safety performance, surpassing comparably sized open-source models and best closed-source commercial models, while maintaining comparable general capabilities.

Problem

• multimodal safety
• adversarial content
• AI safety
• risk detection
• multimodal adversarial robustness

Innovation

• multimodal foundation model
• adversarial robustness
• Confuse-then-Contrast Fine-Tuning
• risk-concept cross-modal alignment
• AI safety