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

Zhiqian Zhang; Xu Zhao; Xiaoqing Xu; Guangdong Liang; Weijia Wang; Xiaolei Lv; Bo Li; Jun Gao

cs.AIMar 31, 2026

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

Authors: Zhiqian Zhang, Xu Zhao, Xiaoqing Xu, Guangdong Liang, Weijia Wang, Xiaolei Lv, Bo Li, Jun Gao

Organizations: Computational Intelligence Dept, Hello Group Inc.

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.

Figures & tables

Content Moderation Safety and compliance judgment Violation attribution	Adversarial Content Defense Adversarial sample recognition Robust defense
Both layers depend on a unified multimodal representation
Content Understanding Provides image-text alignment, OCR perception, spatial reasoning, and general world knowledge representations

Figure 1: Capability layering of the Xuanwu industrial multimodal foundation. The lower layer provides unified multimodal representations, while the upper layers address moderation and adversarial defense.

Figure 2: Core architecture of Xuanwu VL-2B: InternViT-300M serves as the visual backbone and Qwen3 1.7B serves as the language backbone, connected by an MLP projector.

Configurations	InternViT-300M	MLP Projector	Qwen3-1.7B
Architecture Type	ViT-L/14	2-Layer Perceptron	Transformer Decoder
Hidden Size	1024	1024 $\rightarrow$ 2048	2048
Num Heads	16	-	16 (Q) / 8 (KV)
Num Layers	24	2	28
Parameters	$\sim$ 304M	$\sim$ 4M	$\sim$ 1.7B

Table 1: Detailed Hyperparameter Configurations for Xuanwu VL-2B Core Architecture

Stage	Sub-stage	Trainable Modules	Data Type	Scale	Optimization Config
Pre-Training	Modality Alignment	MLP projector	High-quality caption alignment data	1.3M	LR = $1\times 10^{-3}$ , Batch = 256
Pre-Training	General Knowledge Learning	Full Model	General image-text paired data	17.33M	LR = $1\times 10^{-3}$ , Batch = 256
Mid-Training	Business Knowledge Learning	Full Model	Base-retention + instruction-following enhancement + moderation + SPAM adversarial data	2.8M	LR = $2\times 10^{-5}$ , Batch = 128
Post-Training	Instruction Tuning	Full Model	High-fidelity SFT data	8.408M	LR = $1\times 10^{-5}$
Post-Training	RL Alignment	Full Model	GRPO alignment data / adversarial OCR alignment samples	810k	LR = $1\times 10^{-6}$

Table 2: Overview of the Three-Stage, Five-Step Training Recipe

Benchmark	InternVL 3.0 2B	InternVL 3.5 2B	Xuanwu VL-2B
HallusionBench	43.32	46.78	47.32
AI2D	77.75	77.95	82.19
MMStar	58.73	56.20	60.47
OCRBench	81.80	83.10	89.80
MMBench v1.1	79.41	75.08	79.02
MMMU (val)	45.89	50.51	48.11

Table 3: Comparison of General Multimodal Capabilities (%)

Benchmark	InternVL 3.0 2B	InternVL 3.5 2B	Xuanwu VL-2B
IFEval	49.07	69.88	76.35
GPQA Diamond	29.80	34.34	32.83
GSM8K	73.09	74.07	73.92
MMLU	61.74	65.51	62.94
C-Eval	73.12	63.66	62.18
BBH	51.89	62.85	59.68

Table 4: Comparison of Text-Only Capabilities (%)

Category	InternVL 3.0 2B	InternVL 3.5 2B	Xuanwu VL-2B
ad	70.61	88.39	99.38
high-risk	4.50	7.93	97.99
illegal	1.38	12.93	92.19
porn	40.80	61.80	97.15
vulgar	21.63	25.72	79.08
other	4.09	47.50	94.89

Table 5: Business Content Moderation Recall for Independent Binary Decisions (%)

Category	InternVL 3.0 2B	InternVL 3.5 2B	Gemini-2.5-Pro	Xuanwu VL-2B
aigc	16.21	14.34	26.06	32.83
combination	60.60	58.05	74.31	74.62
handwriting	80.99	80.17	89.53	88.86
long	59.87	57.28	80.15	79.22
noise	58.77	55.14	73.27	79.09
small	69.99	77.29	92.82	91.76

Table 6: Business Adversarial OCR Recall Comparison (%)

Metric	InternVL 3.5 2B PT	Xuanwu PT	Xuanwu Mid-Training
Multimodal average-7	58.17	62.29	62.63
Text average-9	58.27	51.96	56.14
Content Moderation average-7	30.04	57.11	94.44
Adversarial OCR weighted overall	64.62	63.75	75.35

Table 7: Stage-wise comparison: InternVL 3.5 2B PT, Xuanwu PT, and Xuanwu Mid-Training (%)

ViT Type	Multimodal	Text	Content Moderation	Adversarial OCR
InternViT-300M	62.63	56.14	94.44	75.35
AIMv2-Huge (600M)	61.88	54.87	96.18	74.91
SAILViT-Huge (600M)	61.68	56.25	95.35	76.73

Table 8: Vision encoder selection comparison after mid-training. Multimodal = average-7; text = average-9; content moderation and adversarial OCR are reported as recall (%).

Stage	Metric	Single	Seq. Fusion	Channel Fusion
Pretrain	HallusionBench	17.84	18.28	17.43
Pretrain	AI2D	36.27	38.60	35.75
Pretrain	MMStar	31.47	30.87	29.50
Pretrain	OCRBench	32.50	32.70	34.50
Pretrain	average-4	29.52	30.11	29.30
SFT	HallusionBench	21.48	26.13	25.04

Table 9: Ablation results for dual-encoder visual fusion. Sequence fusion concatenates the two streams along the visual-token dimension; channel fusion keeps the token count unchanged and concatenates along the feature dimension.

Model Configuration	Weighted Overall (%)	False Positive Rate (FP, %)
Xuanwu VL-2B (Full Pipeline)	82.82	0.08
w/o GRPO	76.42	0.12
w/o SFT Data Curation	79.35	0.31

Table 10: Ablation Studies on Adversarial-Set

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Category	#Datasets	#Samples
Captioning & Knowledge	13	3,956,963
Chart & Table	16	1,968,589
General VQA	34	3,266,107
Grounding & Counting	7	1,464,624
Mathematics	7	567,529
Naive OCR	26	2,232,976

Appendix

Table 11: Category-level statistics of the raw source data for the Pretrain stage

Top-level Category	#Samples	Datasets and References
Captioning & Knowledge	3,956,963	DenseFusion-1M [ 53 ] ; ShareGPT4V [ 54 ] ; CC3M [ 55 ] ; ALLaVA (LAION-Caption) [ 56 ] ; Image-Textualization [ 57 ] ; ALLaVA (VFLAN-Caption) [ 56 ] ; ShareGPT-4o [ 58 ] ; WikiArt [ 59 ] ; Movie-Posters [ 60 ] ; KVQA [ 61 ] ; TextCaps [ 62 ] ; TMDB-Celeb-10K [ 63 ] ; Emo-Visual-Data [ 64 ]
Chart & Table	1,968,589	UniChart [ 65 ] ; PlotQA [ 66 ] ; MMC-Inst [ 67 ] ; DVQA [ 68 ] ; FigureQA [ 69 ] ; Block-Diagram [ 70 ] ; VQAonBD [ 71 ] ; MapQA [ 72 ] ; ChartQA [ 73 ] ; Chart2Text [ 74 ] ; InfoVQA [ 75 ] ; VisText [ 76 ] ; MultiHiertt [ 77 ] ; LRV-Instruction [ 78 ] ; TAT-DQA [ 79 ] ; Diagram-Image-To-Text [ 80 ]
General VQA	3,266,107	ALLaVA (LAION-Instruct, Stage 1.5) [ 56 ] ; ALLaVA (LAION-Instruct, Stage 0.5) [ 56 ] ; MMInstruct-GPT4V [ 81 ] ; LNQA [ 82 ] ; LVIS-Instruct4V [ 83 ] ; ALLaVA (VFLAN-Instruct, Stage 1.5) [ 56 ] ; ALLaVA (VFLAN-Instruct, Stage 0.5) [ 56 ] ; LLaVA-Instruct-150K (ZH) [ 84 ] ; LLaVA-Instruct-150K (EN) [ 84 ] ; NLVR2 [ 85 ] ; RLAIF-V [ 86 ] ; VQAv2 [ 87 ] ; COCO-QA [ 88 ] ; LLaVA-Critic-113K (Pointwise) [ 89 ] ; GQA [ 90 ] ; MIMIC-CGD [ 91 ] ; DaTikZ [ 92 ] ; LLaVA-Critic-113K (Pairwise) [ 89 ] ; KonIQ-10k [ 93 ] ; ICON-QA [ 94 ] ; LLaVAR [ 95 ] ; Places365 [ 96 ] ; A-OKVQA [ 97 ] ; IDK [ 98 ] ; LAION-GPT4V [ 99 ] ; WebSight [ 100 ] ; Hateful Memes [ 12 ] ; Spot-the-Diff [ 101 ] ; Memotion [ 102 ] ; WildVision [ 103 ] ; SketchyVQA [ 104 ] ; VizWiz [ 105 ] ; Indoor Scene Classification [ 106 ] ; DriveLM [ 107 ]
Grounding & Counting	1,464,624	Objects365 [ 108 ] ; TallyQA [ 109 ] ; SA-1B [ 110 ] ; RefCOCO [ 111 ] ; RefCOCO+ [ 111 ] ; SpatialSense [ 112 ] ; GroundUI-18K [ 113 ]
Mathematics	567,529	MAVIS (Function) [ 114 ] ; MAVIS (Geometry) [ 114 ] ; TabMWP [ 115 ] ; CLEVR-Math [ 116 ] ; UniGeo [ 117 ] ; Geometry3K [ 118 ] ; GeoS [ 119 ]
Naive OCR	2,232,976	Latex-Formula [ 120 ] ; SynthDoG-EN [ 121 ] ; SynthDoG-ZH [ 121 ] ; K12-Printing [ 122 ] ; Handwriting-Latex [ 123 ] ; HME-100K [ 124 ] ; TAL-OCR-Composed-37K [ 125 ] ; SROIE [ 126 ] ; LSVT [ 127 ] ; CTW1500 [ 128 ] ; ReCTS [ 129 ] ; COCO-Text [ 130 ] ; Handwritten-Mathematical-Expression [ 131 ] ; CAPTCHA [ 132 ] ; IAM [ 133 ] ; MTWI [ 134 ] ; Chrome-Writting [ 135 ] ; HierText OCR [ 136 ] ; RenderedText [ 137 ] ; IMGUR5K [ 138 ] ; ICDAR 2019 ArT [ 139 ] ; WordArt [ 140 ] ; Invoices and Receipts OCR v1 [ 141 ] ; ORAND-CAR [ 142 ] ; IIIT-5K [ 143 ] ; SVRD [ 144 ]

Appendix

Table 12: Top-level summary of the raw source data for the Pretrain stage and dataset references

Training Stage	Data Block	Scale	Description
Mid-Training	General data	1.898M	1.75M samples are sampled at 10% from the pretraining corpus to preserve the model’s general representations. Another 148k instruction-following enhancement samples, built from AutoIF-Instruct [ 28 ] , IFEval-like Data [ 29 ] , and Tulu-3 Personas [ 30 ] , are used to improve complex instruction understanding and output-format compliance.
Mid-Training	Content moderation	646k	Positive and negative image/text moderation data contribute 598k samples, including 418k image samples and 180k text samples, covering both positive and negative moderation cases. Another 48k OCR and caption annotation samples are used to supplement OCR and caption capability in dynamic-image scenarios.
Mid-Training	Content adversarial	257k	Synthetic SPAM adversarial data contribute 50k samples, primarily built from watermark-distorted OCR images. Human-annotated SPAM adversarial data contribute 40k samples for adversarial recognition. LLM-assisted SPAM adversarial data contribute 167k samples and are all manually reviewed.

Appendix

Table 13: Training data composition of the Mid-Training stage

Training Stage	Data Block	Scale	Description
SFT	General data	8.01M	Composed of 5.14M multimodal samples and 2.87M text-only samples, used to build the foundation for general dialogue, instruction following, OCR/document understanding, and multimodal reasoning.
SFT	Content moderation	359k	Content-safety and moderation-classification data, used to strengthen content safety, moderation classification, and policy consistency.
SFT	Content adversarial	39k	Includes caption, full-image OCR, and violating-text OCR data, used to improve OCR perception and violating-text recognition in real business scenarios.
RL	Content adversarial	810k	Composed of spam-classification and OCR violating-text alignment data, used to further improve task accuracy and robustness in business settings.

Appendix

Table 14: Overall composition of post-training data

Category	Count	Datasets and References
Text & General Instruction	2,303,232	InfinityInstruct [ 169 ] ; AutoIF-Instruct [ 28 ] ; IFEval-like Data [ 29 ] ; Tulu-3 Personas [ 30 ] ; C-Eval (dev split) [ 47 ]
Mathematics	54,888	GSM8K (train split) [ 45 ] ; Orca-Math [ 174 ]
OCR & Document Understanding	494,072	ChartQA [ 73 ] ; DocVQA [ 155 ] ; InfoVQA [ 75 ] ; TextVQA [ 151 ] ; OCR-VQA [ 146 ] ; SROIE [ 126 ] ; POIE [ 184 ] ; FUNSD [ 185 ] ; Imgur5k [ 186 ] ; IAM [ 133 ] ; HME100K [ 124 ] ; Sujet-Finance-QA [ 156 ] ; LAION-COCO [ 187 ] ; DVQA [ 68 ] ; ST-VQA [ 152 ]
Advanced Multimodal Capability	4,645,928	LLaVA-OneVision [ 188 ] ; Cambrian-10M [ 189 ] ; Infinity-MM [ 190 ] ; MMInstruct-GPT4V [ 81 ] ; InternVL2.5 SFT [ 191 ] ; Koniq-10k [ 93 ] ; A-OKVQA [ 97 ] ; M3_Cot [ 192 ]

Appendix

Table 15: Four major categories of the general-data block and representative references

Metric	InternVL PT	Xuanwu PT	Xuanwu Mid-Training
HallusionBench	41.66	47.46	46.92
AI2D	72.02	78.17	77.66
MMStar	47.93	54.73	54.53
OCRBench	78.10	82.90	83.50
MMBench v1.1	70.82	74.46	74.69
MMMU (val)	45.44	44.89	44.00

Appendix

Table 16: Detailed multimodal scores for the stage-wise comparison (%)

Metric	InternVL PT	Xuanwu PT	Xuanwu Mid-Training
IFEval	60.31	46.69	70.50
GPQA Diamond	34.34	21.72	24.75
GSM8K	76.12	77.41	78.92
MMLU	66.77	57.10	63.59
C-Eval	66.81	60.43	60.57
BBH	59.13	58.15	58.73

Appendix

Table 17: Detailed text-only scores for the stage-wise comparison (%)

Category	InternVL PT	Xuanwu PT	Xuanwu Mid-Training
ad	50.05	92.29	98.56
high-risk	7.81	40.95	97.51
illegal	8.60	29.53	98.36
porn	13.03	82.29	98.92
vulgar	10.54	38.23	72.31
other	21.96	17.16	95.91

Appendix

Table 18: Detailed business moderation recall for independent binary decisions in the stage-wise comparison (%)

Category	InternVL PT	Xuanwu PT	Xuanwu Mid-Training
aigc	15.57	22.69	21.66
combination	54.36	63.47	73.45
handwriting	80.26	71.81	88.50
long	52.43	57.93	68.65
noise	54.14	57.75	69.92
small	79.36	77.48	91.39

Appendix

Table 19: Detailed business adversarial OCR scores for the stage-wise comparison (%)

Figure 3: Handwriting-variant case: cursive diversion codes hidden in a complex product image

Figure 4: Layered-watermark case: low-opacity contact information hidden in the image

Figure 5: Micro-font case: miniature variant-character diversion

Figure 6: Text-noise camouflage case: diversion keywords hidden inside long prose paragraphs

Figure 7: Severe distortion and warping case: breaking character structure to evade detection

Figure 8: AIGC-forgery case: contact information fused into generated image textures

Figure 9: Combination-camouflage case: a normal academic background used to hide group numbers

Figure 10: Long-image case: diversion text hidden beneath a routine greeting image

Multimodal Content Moderation

See all 21 papers

Yuvion VL: A Multimodal Foundation Model for Adversarial Content and AI Safety
Jun 23, 2026Shikai Qiu, Xiaowen Xu, Benlei Cui +55VLM Robustness Multimodal Large Language Models
Can Foundation Models Moderate Online Content? Evaluating Instruction- vs. Example-Driven Policy Operationalization
Sep 9, 2026Ayan Majumdar, Shounak Paul, Pushpdeep Singh +6VLM Evaluation LLM Prompting
MPI-SWS, Germany · Saarland University, Germany · INRIA, France
ZeroBench: An Impossible Visual Benchmark for Contemporary Large Multimodal Models
Feb 13, 2025Jonathan Roberts, Mohammad Reza Taesiri, Ansh Sharma +31Visual Reasoning VLM Reasoning
University of Cambridge

Jun 23, 2026J/K move · Enter open · S save

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

Shikai Qiu, Xiaowen Xu, Benlei Cui, Ting Ma+54

VLM Robustness Multimodal Large Language Models LLM Safety Multimodal Model Evaluation

Open paper AskPDF arXiv

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.

Abstract

Multimodal Content Moderation

Explore similar work

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

Can Foundation Models Moderate Online Content? Evaluating Instruction- vs. Example-Driven Policy Operationalization

ZeroBench: An Impossible Visual Benchmark for Contemporary Large Multimodal Models