Xuanwu: Evolving General Multimodal Models into an Industrial-Grade Foundation for Content Ecosystems

2026-03-30

AI Summary

This work addresses the degradation in generalization and catastrophic forgetting observed in current mainstream multimodal models under content moderation and adversarial scenarios, primarily due to insufficient fine-grained visual perception and weak modeling of long-tailed noise. To mitigate these limitations, the authors propose a data-training co-optimization paradigm that integrates a compact architecture—comprising InternViT-300M, an MLP head, and Qwen3-1.7B—with a three-stage progressive training pipeline (pre-training, mid-training, and post-training). This approach effectively balances general-purpose capability retention and domain-specific adaptability within a constrained parameter budget. The resulting model achieves an average score of 67.90 across seven multimodal benchmarks on OpenCompass, an average recall of 94.38% on seven content moderation tasks, and a weighted recall of 82.82% on adversarial OCR-based violation detection, outperforming Gemini-2.5-Pro.

Abstract

In recent years, multimodal large models have continued to improve on general benchmarks. However, in real-world content moderation and adversarial settings, mainstream models still suffer from degraded generalization and catastrophic forgetting because of limited fine-grained visual perception and insufficient modeling of long-tail noise. In this paper, we present Xuanwu VL-2B as a case study of how general multimodal models can be developed into an industrial-grade foundation model for content ecosystems. The model adopts a compact InternViT-300M + MLP + Qwen3 1.7B architecture, balancing fine-grained visual perception, language-semantic alignment, and deployment cost within an approximately 2B-parameter budget. To balance business specialization with the retention of general capabilities, we developed a data iteration and curation mechanism and trained the model through a progressive three-stage pipeline: pre-training, mid-training, and post-training. Ablation studies and offline business evaluations show that Xuanwu VL-2B achieves an average score of 67.90 across seven OpenCompass multimodal metrics (vs. 64.27 for InternVL 3.5 2B), an average recall of 94.38% over seven independent business moderation tasks, and a weighted overall recall of 82.82% on policy-violating text in challenging adversarial OCR scenarios, outperforming Gemini-2.5-Pro (76.72%). These results show that, under a limited parameter budget, Xuanwu VL-2B achieves a practical balance among business alignment, visual perception, general capability retention, and deployment cost.

Problem

• multimodal models
• content moderation
• catastrophic forgetting
• fine-grained visual perception
• long-tail noise

Innovation

• industrial-grade foundation model
• multimodal alignment
• progressive training pipeline
• fine-grained visual perception
• adversarial content moderation