pith. sign in

arxiv: 2606.25034 · v1 · pith:EZZHPBLUnew · submitted 2026-06-23 · 💻 cs.CV · cs.AI

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

Shikai Qiu , Xiaowen Xu , Benlei Cui , Ting Ma , Xiufeng Huang , Wenjing Jiang , Shaoxuan He , Haolei Xu
show 46 more authors
Chunyang Chai Yujian Li Yiliang Zhang Guanghui Wang Ziheng Wang Ziwen Xu Zhaoyu Fan Jinhao Chen Ruijie Jian Hongxing Li Chuxi Xiao Xinyue Chen Wenxuan Liu Libin Dong Yupeng Cao Xiaoqian Xia Jing Wang Zhe Jiang Zhenan Ye Guang Yang Bin Liu Wei Peng Ziqiang Zhu Meihui Lian Kaiwen Lv Kacuila Haidong Ding Dongjie Zhang Yangfan Zhou Bingyu Zhu Yan Wang Hai Zhao Xuan Jin Wei Zhao Pengfei Sun Huiming Zhang Wei Wang Xipeng Cao Bin Li Chengwen Yao Meng Huang Xianfeng Li Bin Tang Chao Liu Hui Xue Longtao Huang Haiwen Hong
This is my paper

Pith reviewed 2026-06-26 00:01 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords multimodal large language modelsAI safetyadversarial robustnesscontent moderationcontrastive fine-tuningrisk evaluation benchmarksinstruction tuningreasoning models
0
0 comments X

The pith

Yuvion VL-32B achieves top safety performance on multimodal adversarial tasks while matching general capabilities of other models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Yuvion VL as a family of multimodal large language models built specifically to handle content and AI safety risks that arise from adversarial, cross-modal inputs. It constructs training data through an automated adversarial-aware synthesis process and applies a three-stage pipeline of continued pretraining, instruction tuning, and reasoning post-training, plus a new Confuse-then-Contrast Fine-Tuning step that forces the model to separate visually similar inputs with different safety meanings. The central result is that the 32B variant outperforms both open-source models of similar size and leading closed-source systems on the authors' YVRE safety benchmarks without loss in general tasks. A reader would care because most general-purpose multimodal models still fail to reliably flag real-world safety threats that combine images and text in deceptive ways.

Core claim

Yuvion VL treats safety as an inherently adversarial and multimodal problem and designs its full pipeline around robustness. An automated data pipeline produces large-scale multimodal samples with domain knowledge and reasoning annotations. Training proceeds in three stages: continued pretraining for risk-concept alignment, instruct post-training for production safety tasks, and reasoning post-training for interpretability. The key addition is Confuse-then-Contrast Fine-Tuning, which mines model confusions and builds multi-image contrastive groups to sharpen discrimination of fine-grained visual-semantic differences. On the introduced YVRE benchmark collection the 32B model surpasses compara

What carries the argument

Confuse-then-Contrast Fine-Tuning: a contrastive method that identifies model-specific confusions and builds multi-image groups to enforce explicit separation of visually similar cases that carry different safety implications.

If this is right

  • Production systems can deploy a single multimodal model that meets high safety standards without separate moderation layers.
  • Reasoning-oriented variants improve interpretability of safety decisions in complex cases.
  • The three-stage pipeline plus contrastive fine-tuning scales to other multimodal safety applications.
  • YVRE supplies a standardized benchmark for comparing future adversarial-robust safety models.
  • Models trained this way maintain general capabilities, reducing the usual trade-off between safety and utility.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the performance holds on unseen adversarial patterns, organizations may shift from closed-source safety APIs to open fine-tuned models.
  • The contrastive approach could extend to other domains requiring fine discrimination, such as medical image analysis or autonomous vehicle perception.
  • Independent reproduction of YVRE would be needed before regulators treat the reported safety margins as reliable.
  • Larger-scale versions of the same pipeline might further close the gap with frontier closed models on safety without capability loss.

Load-bearing premise

The Yuvion VL RiskEval collection and the authors' internal evaluations provide an unbiased and comprehensive measure of real-world adversarial robustness.

What would settle it

An independent test set of adversarial multimodal examples, constructed without access to the authors' data or model, on which Yuvion VL-32B scores below the best competing models on safety metrics.

Figures

Figures reproduced from arXiv: 2606.25034 by Benlei Cui, Bingyu Zhu, Bin Li, Bin Liu, Bin Tang, Chao Liu, Chengwen Yao, Chunyang Chai, Chuxi Xiao, Dongjie Zhang, Guanghui Wang, Guang Yang, Haidong Ding, Haiwen Hong, Hai Zhao, Haolei Xu, Hongxing Li, Huiming Zhang, Hui Xue, Jing Wang, Jinhao Chen, Kaiwen Lv Kacuila, Libin Dong, Longtao Huang, Meihui Lian, Meng Huang, Pengfei Sun, Ruijie Jian, Shaoxuan He, Shikai Qiu, Ting Ma, Wei Peng, Wei Wang, Wei Zhao, Wenjing Jiang, Wenxuan Liu, Xianfeng Li, Xiaoqian Xia, Xiaowen Xu, Xinyue Chen, Xipeng Cao, Xiufeng Huang, Xuan Jin, Yangfan Zhou, Yan Wang, Yiliang Zhang, Yujian Li, Yupeng Cao, Zhaoyu Fan, Zhe Jiang, Zhenan Ye, Ziheng Wang, Ziqiang Zhu, Ziwen Xu.

Figure 1
Figure 1. Figure 1: Performance comparison between Yuvion VL and other models on open-source safety bench [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the automated Visual CoT production and quality-inspection pipeline for risk [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the Yuvion VL training pipeline. The pipeline consists of three stages: Continued [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the C2FT framework. (a) Dynamic construction of a semantic confusion set [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the training pipeline for Yuvion VL Reasoning model. The pipeline consists [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Framework of the Yuvion VL RiskEval (YVRE). [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative case studies of Yuvion VL across eight risk scenarios. For each case, we compare [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper presents Yuvion VL, a family of multimodal LLMs purpose-built for content and AI safety. It describes an automated adversarial-aware data synthesis pipeline with multi-stage quality control, a three-stage training process (continued pretraining for cross-modal risk alignment, instruct post-training, and reasoning post-training), and introduces Confuse-then-Contrast Fine-Tuning to mine model confusions and enforce discrimination via multi-image contrastive groups. The authors further introduce the Yuvion VL RiskEval (YVRE) benchmark collection covering open and internal safety, adversarial robustness, and capability evaluations, claiming that the 32B variant achieves industry-leading safety performance that surpasses comparably sized open-source models and best closed-source commercial models while maintaining comparable general capabilities.

Significance. If the performance claims can be substantiated on independent benchmarks, the work would advance multimodal safety modeling by treating adversarial robustness as a core design principle and introducing a contrastive fine-tuning method that targets fine-grained visual-semantic distinctions. The three-stage pipeline and data synthesis approach provide concrete engineering contributions that could be adopted more broadly. The significance is currently limited by the evaluation design.

major comments (1)
  1. [YVRE introduction and experiments section] The headline claim that Yuvion VL-32B achieves industry-leading safety performance (abstract and experiments) rests entirely on results from the newly introduced YVRE collection, which the authors control in both construction and selection. No comparisons are reported against established public multimodal safety benchmarks such as MM-SafetyBench or SafeBench, and no third-party red-teaming results are provided. This is load-bearing because the central empirical contribution cannot be assessed for selection bias, data leakage from the three-stage pipeline plus Confuse-then-Contrast procedure, or over-representation of synthesized adversarial patterns without external validation.
minor comments (1)
  1. [Abstract] The abstract asserts superior performance without any quantitative metrics, baselines, or error bars; a brief summary of key numbers should be included even in the abstract for a results-oriented claim.

Simulated Author's Rebuttal

1 responses · 1 unresolved

We thank the referee for highlighting the evaluation design limitations. We address the concern about reliance on the internally controlled YVRE benchmarks below and commit to revisions that incorporate external validation where feasible.

read point-by-point responses

  1. Referee: [YVRE introduction and experiments section] The headline claim that Yuvion VL-32B achieves industry-leading safety performance (abstract and experiments) rests entirely on results from the newly introduced YVRE collection, which the authors control in both construction and selection. No comparisons are reported against established public multimodal safety benchmarks such as MM-SafetyBench or SafeBench, and no third-party red-teaming results are provided. This is load-bearing because the central empirical contribution cannot be assessed for selection bias, data leakage from the three-stage pipeline plus Confuse-then-Contrast procedure, or over-representation of synthesized adversarial patterns without external validation.

    Authors: We agree that the primary safety claims are evaluated on YVRE and that external benchmarks are needed to assess potential selection bias or leakage from our data synthesis and Confuse-then-Contrast procedure. YVRE targets fine-grained adversarial multimodal distinctions not comprehensively covered by MM-SafetyBench or SafeBench, which is why it was introduced. To strengthen the claims, the revised manuscript will add performance comparisons on both MM-SafetyBench and SafeBench. We do not currently possess third-party red-teaming results and cannot generate them internally; instead, we will release model weights and the YVRE suite to support independent verification. revision: partial

standing simulated objections not resolved
  • Third-party red-teaming results (no access to external evaluations)

Circularity Check

0 steps flagged

No circularity in derivation chain; empirical claims rest on introduced benchmark with open components

full rationale

The paper describes a three-stage training pipeline and Confuse-then-Contrast method for a safety-focused multimodal model, then reports performance on the newly introduced YVRE benchmark collection (which explicitly includes open evaluations alongside internal ones). No equations, fitted parameters, or first-principles derivations are presented that reduce by construction to the inputs; the central empirical claim of superior safety performance is a comparison result on the benchmark rather than a self-definitional or tautological reduction. Self-created benchmarks are common and do not trigger the enumerated circularity patterns unless data leakage or direct renaming of training outputs as predictions is shown, which is not exhibited here. The derivation chain remains self-contained against external benchmarks and comparisons.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no information on free parameters, background axioms, or new postulated entities is provided.

pith-pipeline@v0.9.1-grok · 6022 in / 1125 out tokens · 38230 ms · 2026-06-26T00:01:01.967560+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 17 linked inside Pith

  1. [1]

    Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966,

    Jinze Bai, Shuai Bai, Shusheng Yang, et al. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966,

    Pith/arXiv arXiv

  2. [2]

    Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025a

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, et al. Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025a. Shuai Bai, Keqin Chen, Xuejing Liu, et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025b. Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, ...

    Pith/arXiv arXiv

  3. [3]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks.arXiv preprint arXiv:2312.14238,

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks.arXiv preprint arXiv:2312.14238,

    Pith/arXiv arXiv

  4. [4]

    Llama guard 3 vision: Safeguarding human-ai image understanding conversations.arXiv preprint arXiv:2411.10414,

    Jianfeng Chi, Ujjwal Karn, Hongyuan Zhan, Eric Smith, Javier Rando, Yiming Zhang, Kate Plawiak, Zacharie Delpierre Coudert, Kartikeya Upasani, and Mahesh Pasupuleti. Llama guard 3 vision: Safeguarding human-ai image understanding conversations.arXiv preprint arXiv:2411.10414,

    arXiv

  5. [5]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pp. 2924–2936,

    2019

  6. [6]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Pith/arXiv arXiv

  7. [7]

    Tc-pad\’e: Trajectory-consistent pad\’e approximation for diffusion acceleration.arXiv preprint arXiv:2603.02943, 2026a

    Benlei Cui, Shaoxuan He, Bukun Huang, Zhizeng Ye, Yunyun Sun, Longtao Huang, Hui Xue, Yang Yang, Jingqun Tang, Zhou Zhao, et al. Tc-pad\’e: Trajectory-consistent pad\’e approximation for diffusion acceleration.arXiv preprint arXiv:2603.02943, 2026a. 1Correspondence to:honghaiwen.hhw@alibaba-inc.com. 21 Benlei Cui, Bukun Huang, Zhizeng Ye, Xuemei Dong, Tuo...

    arXiv

  8. [8]

    AEGIS: Online adaptive ai content safety moderation with ensemble of llm experts.arXiv preprint arXiv:2404.05993,

    Shaona Ghosh, Prasoon Varshney, Erick Galinkin, and Christopher Parisien. AEGIS: Online adaptive ai content safety moderation with ensemble of llm experts.arXiv preprint arXiv:2404.05993,

    arXiv

  9. [9]

    Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms.arXiv preprint arXiv:2406.18495,

    Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms.arXiv preprint arXiv:2406.18495,

    Pith/arXiv arXiv

  10. [10]

    Llava- Guard: An open vlm-based framework for safeguarding vision datasets and models.arXiv preprint arXiv:2406.05113,

    Lukas Helff, Felix Friedrich, Manuel Brack, Kristian Kersting, and Patrick Schramowski. Llava- Guard: An open vlm-based framework for safeguarding vision datasets and models.arXiv preprint arXiv:2406.05113,

    arXiv

  11. [11]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749,

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, et al. Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749,

    Pith/arXiv arXiv

  12. [12]

    Llama guard: Llm-based input-output safeguard for human-ai conversations.arXiv preprint arXiv:2312.06674,

    Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations.arXiv preprint arXiv:2312.06674,

    Pith/arXiv arXiv

  13. [13]

    Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326,

    Bo Li, Yuanhan Zhang, Dong Guo, et al. Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326,

    Pith/arXiv arXiv

  14. [14]

    Seed-bench: Benchmarking multimodal llms with generative comprehension.arXiv preprint arXiv:2307.16125,

    22 Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension.arXiv preprint arXiv:2307.16125,

    Pith/arXiv arXiv

  15. [15]

    Yufeng-xguard: A reasoning-centric, interpretable, and flexible guardrail model for large language models.arXiv preprint arXiv:2601.15588,

    Junyu Lin, Meizhen Liu, Xiufeng Huang, Jinfeng Li, Haiwen Hong, Xiaohan Yuan, Yuefeng Chen, Longtao Huang, Hui Xue, Ranjie Duan, et al. Yufeng-xguard: A reasoning-centric, interpretable, and flexible guardrail model for large language models.arXiv preprint arXiv:2601.15588,

    arXiv

  16. [16]

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. InInternational Conference on Learning Representations, volume 2024, pp. 23439–23554,

    2024

  17. [17]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530,

    Machel Reid, Nikolay Savinov, Denis Teplyashin, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530,

    Pith/arXiv arXiv

  18. [18]

    Gemini: A family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    Gemini Team, Rohan Anil, Sebastian Borgeaud, et al. Gemini: A family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    Pith/arXiv arXiv

  19. [19]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191,

    Peng Wang, Shuai Bai, Sinan Tan, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191,

    Pith/arXiv arXiv

  20. [20]

    Llava-critic-r1: Your critic model is secretly a strong policy model.arXiv preprint arXiv:2509.00676, 2025a

    Xiyao Wang et al. Llava-critic-r1: Your critic model is secretly a strong policy model.arXiv preprint arXiv:2509.00676, 2025a. 23 Yaoting Wang, Shengqiong Wu, Yuecheng Zhang, Shuicheng Yan, Ziwei Liu, Jiebo Luo, and Hao Fei. Multimodal chain-of-thought reasoning: A comprehensive survey.arXiv preprint arXiv:2503.12605, 2025b. Haoning Wu, Zicheng Zhang, Erl...

    arXiv 2024

  21. [21]

    EVADE- Bench: Multimodal benchmark for evaluating and enhancing evasive content detection.arXiv preprint arXiv:2505.17654,

    Ancheng Xu, Zhihao Yang, Jingpeng Li, Guanghu Yuan, Longze Chen, Liang Yan, et al. EVADE- Bench: Multimodal benchmark for evaluating and enhancing evasive content detection.arXiv preprint arXiv:2505.17654,

    Pith/arXiv arXiv

  22. [22]

    Llava-cot: Let vision language models reason step-by-step.arXiv preprint arXiv:2411.10440,

    Guowei Xu, Peng Jin, Ziang Wu, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. Llava-cot: Let vision language models reason step-by-step.arXiv preprint arXiv:2411.10440,

    Pith/arXiv arXiv

  23. [23]

    Seeing but not thinking: Routing distraction in multimodal mixture-of-experts.arXiv preprint arXiv:2604.08541,

    Haolei Xu, Haiwen Hong, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang, Hui Xue, Yongliang Shen, Weiming Lu, and Yueting Zhuang. Seeing but not thinking: Routing distraction in multimodal mixture-of-experts.arXiv preprint arXiv:2604.08541,

    Pith/arXiv arXiv

  24. [24]

    ProGuard: Towards proactive multimodal safeguard.arXiv preprint arXiv:2512.23573,

    Shaohan Yu, Lijun Li, Chenyang Si, Lu Sheng, and Jing Shao. ProGuard: Towards proactive multimodal safeguard.arXiv preprint arXiv:2512.23573,

    arXiv

  25. [25]

    Shieldgemma: Generative ai content moderation based on gemma.arXiv preprint arXiv:2407.21772,

    Wenjun Zeng, Yuchi Liu, Ryan Mullins, Ludovic Peran, Joe Fernandez, Hamza Harkous, Karthik Narasimhan, Drew Proud, Piyush Kumar, Bhaktipriya Radharapu, Olivia Sturman, and Oscar Wahltinez. Shieldgemma: Generative ai content moderation based on gemma.arXiv preprint arXiv:2407.21772,

    Pith/arXiv arXiv

  26. [26]

    R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization.arXiv preprint arXiv:2503.12937, 2025a

    Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization.arXiv preprint arXiv:2503.12937, 2025a. Zicheng Zhang, Haoning Wu, Chunyi Li, Yingjie Zhou, Wei Sun, Xiongkuo Min, Zijian Chen, Xiaohong Liu, Weisi ...

    Pith/arXiv arXiv 2025

  27. [27]

    for diagram understanding and struc- tured visual reasoning over educational figures. • MM General VQA:MMBench(Liu et al., 2024c),MME(Fu et al., 2026),MMStar(Chen et al., 2024),SEEDBench(Li et al., 2023),ScienceQA(Lu et al., 2022),A-Bench(Zhang et al., 2025b), and Q-Bench(Wu et al.,

    2026

  28. [28]

    • Text Chinese language understanding.:C3(Sun et al., 2020),CLUEWSC(Xu et al., 2020), andXiezhi- CNfor Chinese knowledge understanding, and commonsense reasoning

    for object counting and spatial grounding. • Text Chinese language understanding.:C3(Sun et al., 2020),CLUEWSC(Xu et al., 2020), andXiezhi- CNfor Chinese knowledge understanding, and commonsense reasoning. • Text Commonsense and reading comprehension. BoolQ(Clark et al., 2019),WinoGrande(Sakaguchi et al.,

    2020

  29. [29]

    for commonsense reasoning, reading comprehension, multi-step inference, and robustness on open-ended or ambiguous problems. • Text Mathematical reasoning.This group includesGSM8K-ZH(Cobbe et al., 2021),APE210K,TAL- SCQ5K-CNfor arithmetic problem solving, formal mathematical reasoning, and theorem-related question answering in both Chinese and English. • T...

    2021