pith. machine review for the scientific record. sign in

arxiv: 2604.21409 · v1 · submitted 2026-04-23 · 💻 cs.CV

Recognition: unknown

S1-VL: Scientific Multimodal Reasoning Model with Thinking-with-Images

Lifeng Xu, Mingwei Ou, Nan Xu, QingLi Wang, Qingxiao Li, Shu Hu, Yudong Bai

Pith reviewed 2026-05-09 22:23 UTC · model grok-4.3

classification 💻 cs.CV
keywords multimodal reasoningthinking with imagesscientific AIimage processing codedata filteringreinforcement learningchain of thought
0
0 comments X

The pith

S1-VL enables models to reason about scientific images by generating and executing Python code to manipulate them iteratively.

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

The paper introduces S1-VL to address limitations in current multimodal models for scientific tasks by supporting two modes: structured chain-of-thought reasoning and Thinking-with-Images, where the model actively edits visuals via code. Training data comes from six disciplines and passes through a quality filter plus an adaptive router that shifts low-gain visual samples to text-only reasoning. A four-stage process of supervised fine-tuning and reinforcement learning produces the final model. If this works, it would allow AI systems to handle high-resolution charts, microscope images, and geometry problems more reliably than text-only approaches.

Core claim

S1-VL natively combines Scientific Reasoning with Thinking-with-Images, in which the model outputs Python code for image operations, runs it in a sandbox to receive intermediate visual results, and continues multi-turn reasoning; the 32B version trained this way reaches state-of-the-art results on all five Thinking-with-Images benchmarks and leads on scientific reasoning sets such as Physics and VRSBench.

What carries the argument

The Thinking-with-Images mode together with the adaptive data routing strategy, which converts samples that yield little visual gain into pure Reasoning-mode data so the model learns when code-based image operations are actually required.

If this is right

  • The model outperforms prior systems on HRBench-4K, HRBench-8K, MME-RealWorld-CN, MME-RealWorld-Lite, and V*.
  • It also leads on scientific reasoning benchmarks such as Physics and VRSBench.
  • Training teaches the model to choose pure reasoning for some inputs instead of always invoking image operations.
  • The progressive four-stage pipeline of scientific SFT, cold-start Thinking-with-Images SFT, and two RL stages produces the observed gains.

Where Pith is reading between the lines

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

  • This code-in-the-loop design could transfer to other visual domains that need verifiable intermediate steps, such as medical image analysis.
  • It implies that hybrid symbolic and visual reasoning may be more effective than either alone when images contain precise quantitative information.
  • Future tests could measure whether the routing strategy reduces unnecessary computation on simple visual questions.

Load-bearing premise

The six-dimensional quality filtering framework and adaptive routing strategy correctly identify and repurpose low visual-information-gain samples without discarding useful signal or introducing bias.

What would settle it

Run S1-VL on a held-out set of scientific images that require exact pixel measurements or geometric transformations only possible through code, and check whether accuracy drops below text-only baselines or prior multimodal systems.

Figures

Figures reproduced from arXiv: 2604.21409 by Lifeng Xu, Mingwei Ou, Nan Xu, QingLi Wang, Qingxiao Li, Shu Hu, Yudong Bai.

Figure 1
Figure 1. Figure 1: Benchmark performance of S1-VL-32B. 1 arXiv:2604.21409v1 [cs.CV] 23 Apr 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Thinking-with-Images multi-turn inference pipeline in S1-VL. Given an input [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the two parallel data processing pipelines for scientific reasoning and Thinking [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The four-stage progressive training pipeline of S1-VL. Each stage builds upon the previous one, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of reward trajectories before and after reward function revision. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Thinking-with-Images case on a radiology CT image (medical domain). The model crops and [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Thinking-with-Images case on a remote sensing image from VRSBench (geography domain). [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Thinking-with-Images case on TEM diffraction pattern analysis (chemistry/materials science [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Scientific reasoning case on a multi-image mechanics problem (physics domain). The model [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Scientific reasoning case on a number-strip problem (mathematics domain). The model extracts [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Scientific reasoning case on a galaxy morphology classification task. The model identifies the [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Failure case: imprecise spatial grounding with self-correction. The first crop captures an [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Failure case: spurious success via language priors. The cropped region misses the red-rectangle [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
read the original abstract

We present S1-VL, a multimodal reasoning model for scientific domains that natively supports two complementary reasoning paradigms: Scientific Reasoning, which relies on structured chain-of-thought, and Thinking-with-Images, which enables the model to actively manipulate images through Python code execution during reasoning. In the Thinking-with-Images mode, the model generates and executes image-processing code in a sandbox environment, obtains intermediate visual results, and continues reasoning in a multi-turn iterative manner. This design is particularly effective for challenging scenarios such as high-resolution scientific chart interpretation, microscopic image understanding, and geometry-assisted reasoning. To construct the training data, we collect scientific multimodal datasets spanning six disciplines: mathematics, physics, chemistry, astronomy, geography, and biology. We further develop a six-dimensional quality filtering framework for reasoning trajectories. To mitigate redundant, ineffective, and erroneous visual operations commonly found in existing datasets, we propose a multi-stage filtering pipeline together with an adaptive data routing strategy. This strategy converts samples with low visual information gain into pure Reasoning-mode data, enabling the model to learn when image operations are truly necessary. S1-VL is trained through a four-stage progressive pipeline: scientific multimodal SFT, Thinking-with-Images cold-start SFT, and two stages of reinforcement learning with SAPO. We build S1-VL-32B on top of Qwen3-VL-32B-Thinking and evaluate it on 13 benchmarks. Experimental results show that S1-VL-32B achieves state-of-the-art performance on all five Thinking-with-Images benchmarks, including HRBench-4K, HRBench-8K, MME-RealWorld-CN, MME-RealWorld-Lite, and V*, and outperforms compared systems on scientific reasoning benchmarks such as Physics and VRSBench.

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

2 major / 1 minor

Summary. The paper introduces S1-VL, a 32B multimodal model for scientific domains supporting two reasoning modes: structured chain-of-thought (Scientific Reasoning) and Thinking-with-Images, in which the model generates/executes Python code for image manipulation in a sandbox to enable iterative visual reasoning on tasks like high-resolution charts and geometry. Data are collected across six disciplines (math, physics, chemistry, astronomy, geography, biology); a six-dimensional quality filter plus multi-stage pipeline and adaptive routing convert low visual-information-gain samples to pure text reasoning. Training follows a four-stage pipeline (scientific multimodal SFT, Thinking-with-Images cold-start SFT, two SAPO RL stages) starting from Qwen3-VL-32B-Thinking. On 13 benchmarks, S1-VL-32B claims SOTA on all five Thinking-with-Images tasks (HRBench-4K/8K, MME-RealWorld-CN/Lite, V*) and outperforms baselines on scientific reasoning benchmarks such as Physics and VRSBench.

Significance. If the performance deltas are shown to be robust, the work would advance multimodal scientific reasoning by demonstrating that explicit code-based visual manipulation can be learned and routed effectively, addressing a gap in handling complex visual scientific data. The adaptive routing mechanism to avoid ineffective image operations and the progressive SFT-to-RL pipeline are practical contributions that could generalize. The approach of turning low-gain samples into reasoning-only data is a sensible engineering response to training noise, though its net benefit remains unquantified.

major comments (2)
  1. [Data Construction section] Data Construction section: The six-dimensional quality filtering framework and adaptive data routing strategy are presented as central to producing effective training data and mitigating ineffective visual operations, yet the manuscript supplies no ablation studies that isolate their effect on the final benchmark suite. Without controlled comparisons (e.g., training the same backbone with vs. without the routing threshold on HRBench-4K, MME-RealWorld variants, and V*), it is impossible to determine whether the reported SOTA margins arise from the Thinking-with-Images cold-start and SAPO RL stages or from data selection alone. This directly undermines attribution of the central performance claims.
  2. [Experimental Results section] Experimental Results section: The claims of SOTA performance on all five Thinking-with-Images benchmarks and outperformance on Physics/VRSBench are stated without accompanying details on the exact baselines, statistical significance tests, ablation results for the four training stages, or error analysis. In the absence of these controls, the robustness of the gains versus Qwen3-VL-32B-Thinking cannot be assessed, leaving the effectiveness of the overall pipeline unverified.
minor comments (1)
  1. [Data Construction section] The six dimensions of the quality filtering framework are referenced but not enumerated with explicit criteria or a summary table; adding this would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We agree that the manuscript would benefit from additional ablation studies and expanded experimental details to strengthen attribution of results and verify robustness. We address each major comment below and commit to revisions accordingly.

read point-by-point responses

  1. Referee: [Data Construction section] Data Construction section: The six-dimensional quality filtering framework and adaptive data routing strategy are presented as central to producing effective training data and mitigating ineffective visual operations, yet the manuscript supplies no ablation studies that isolate their effect on the final benchmark suite. Without controlled comparisons (e.g., training the same backbone with vs. without the routing threshold on HRBench-4K, MME-RealWorld variants, and V*), it is impossible to determine whether the reported SOTA margins arise from the Thinking-with-Images cold-start and SAPO RL stages or from data selection alone. This directly undermines attribution of the central performance claims.

    Authors: We appreciate the referee highlighting this gap in experimental validation. The manuscript describes the six-dimensional quality filtering and adaptive routing but does not include isolating ablations. We will add controlled comparisons in the revised manuscript: training the same backbone with versus without the routing threshold, evaluated specifically on HRBench-4K, MME-RealWorld-CN, MME-RealWorld-Lite, and V*. These results will quantify the contribution of data selection relative to the Thinking-with-Images cold-start SFT and SAPO RL stages. revision: yes

  2. Referee: [Experimental Results section] Experimental Results section: The claims of SOTA performance on all five Thinking-with-Images benchmarks and outperformance on Physics/VRSBench are stated without accompanying details on the exact baselines, statistical significance tests, ablation results for the four training stages, or error analysis. In the absence of these controls, the robustness of the gains versus Qwen3-VL-32B-Thinking cannot be assessed, leaving the effectiveness of the overall pipeline unverified.

    Authors: We acknowledge that the current Experimental Results section reports aggregate SOTA and outperformance claims without the full set of requested controls. In the revision we will expand this section to: explicitly list all baseline models with configurations; report statistical significance tests (e.g., paired t-tests or bootstrap p-values) for gains over Qwen3-VL-32B-Thinking; provide stage-wise ablations for the four training stages; and include error analysis on representative failure cases. These additions will allow direct assessment of pipeline robustness. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain or performance claims

full rationale

The paper's central claims consist of empirical performance gains on held-out external benchmarks (HRBench-4K/8K, MME-RealWorld variants, V*, Physics, VRSBench) after a described four-stage training pipeline on collected scientific multimodal data. The six-dimensional filtering and adaptive routing are methodological choices for data preparation, not quantities that the final results reduce to by definition or construction. No equations, self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The derivation chain (data collection → filtering/routing → SFT/RL stages → evaluation) remains self-contained against independent test sets, with no reduction of outputs to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unverified effectiveness of the multi-stage filtering pipeline, adaptive routing, and four-stage training process; these are presented as novel contributions but lack independent validation details in the abstract.

axioms (1)
  • domain assumption Standard multimodal SFT and RL training stages improve performance when applied to the described data and modes.
    The four-stage progressive pipeline assumes these established methods transfer effectively to the new Thinking-with-Images capability.

pith-pipeline@v0.9.0 · 5646 in / 1303 out tokens · 78019 ms · 2026-05-09T22:23:15.124826+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

24 extracted references · 22 canonical work pages · 11 internal anchors

  1. [1]

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities.arXiv:2308.12966,

    work page internal anchor Pith review arXiv

  2. [2]

    Intern-s1: A scientific multimodal foundation model.arXiv preprint arXiv:2508.15763, 2025

    Lei Bai, Zhongrui Cai, Yuhang Cao, Maosong Cao, Weihan Cao, Chiyu Chen, Haojiong Chen, Kai Chen, Pengcheng Chen, Ying Chen, et al. Intern-s1: A scientific multimodal foundation model.arXiv preprint arXiv:2508.15763, 2025a. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Q...

    work page arXiv

  3. [3]

    InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks

    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,

    work page internal anchor Pith review arXiv

  4. [4]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261,

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Open- vlthinker: Complex vision-language reasoning via iterative sft-rl cycles.arXiv preprint arXiv:2503.17352, 2025

    Yihe Deng et al. Openvlthinker: An early exploration to complex vision-language reasoning via iterative self-refinement.arXiv preprint arXiv:2503.17352,

    work page arXiv

  6. [6]

    Physics: Benchmarking foundation models on university-level physics problem solving

    Kaiyue Feng, Yilun Zhao, Yixin Liu, Tianyu Yang, Chen Zhao, John Sous, and Arman Cohan. Physics: Benchmarking foundation models on university-level physics problem solving. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 11717–11743,

    2025

  7. [7]

    CoRR , volume =

    Chang Gao, Chujie Zheng, Xiong-Hui Chen, Kai Dang, Shixuan Liu, Bowen Yu, An Yang, Shuai Bai, Jingren Zhou, and Junyang Lin. Soft adaptive policy optimization.arXiv preprint arXiv:2511.20347,

    work page arXiv

  8. [8]

    19 Tiechuan Hu, Wenbo Zhu, and Yuqi Yan

    URLhttps://github.com/henrysky/Galaxy10. 19 Tiechuan Hu, Wenbo Zhu, and Yuqi Yan. Reward hacking in reinforcement learning and rlhf: A multidisciplinary examination of vulnerabilities, mitigation strategies, and alignment challenges. In 2025 5th Intelligent Cybersecurity Conference (ICSC), pp. 272–275. IEEE,

    2025

  9. [9]

    Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Xu Tang, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749,

    work page internal anchor Pith review arXiv

  10. [10]

    Visual agentic reinforcement fine-tuning

    Ziyu Liu, Yuhang Zang, Yushan Zou, Zijian Liang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual agentic reinforcement fine-tuning.arXiv preprint arXiv:2505.14246,

    work page arXiv

  11. [11]

    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.arXiv preprint arXiv:2310.02255,

    work page internal anchor Pith review arXiv

  12. [12]

    MM-Eureka: Exploring the Frontiers of Multimodal Reasoning with Rule-based Reinforcement Learning

    Fanqing Meng et al. Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning.arXiv preprint arXiv:2503.07365,

    work page Pith review arXiv

  13. [13]

    URL https://openai.com/index/ introducing-openai-o1-preview/. OpenAI. Gpt-5, 2025a. URLhttps://openai.com/zh-Hans-CN/index/introducing-gpt-5/. OpenAI. Introducing openai o3 and o4-mini, 2025b. URL https://openai.com/zh-Hans-CN/index/ introducing-o3-and-o4-mini/. OpenAI. Thinking with images, 2025c. URLhttps://openai.com/index/thinking-with-images/. Chris ...

    work page arXiv

  14. [14]

    V-thinker: Interactive thinking with images.arXiv preprint arXiv:2511.04460, 2025

    Runqi Qiao, Qiuna Tan, Minghan Yang, Guanting Dong, Peiqing Yang, Shiqiang Lang, Enhui Wan, Xiaowan Wang, Yida Xu, Lan Yang, et al. V-thinker: Interactive thinking with images.arXiv preprint arXiv:2511.04460,

    work page arXiv

  15. [15]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Thinking with Images for Multimodal Reasoning: Foundations, Methods, and Future Frontiers

    Zhaochen Su, Peng Xia, Hangyu Guo, Zhenhua Liu, Yan Ma, Xiaoye Qu, Jiaqi Liu, Yanshu Li, Kaide Zeng, Zhengyuan Yang, et al. Thinking with images for multimodal reasoning: Foundations, methods, and future frontiers.arXiv preprint arXiv:2506.23918, 2025a. Zhaochen Su et al. Openthinkimg: Learning to think with images via visual tool reinforcement learning. ...

    work page internal anchor Pith review arXiv

  17. [17]

    Gemini: A Family of Highly Capable Multimodal Models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning

    Haoming Wang, Haoyang Zou, Huatong Song, Jiazhan Feng, Junjie Fang, Junting Lu, Longxiang Liu, Qinyu Luo, Shihao Liang, Shijue Huang, et al. Ui-tars-2 technical report: Advancing gui agent with multi-turn reinforcement learning.arXiv preprint arXiv:2509.02544, 2025a. 20 Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and...

    work page internal anchor Pith review arXiv

  19. [19]

    Zooming without zooming: Region-to-image distillation for fine-grained multimodal perception.arXiv preprint arXiv:2602.11858, 2026

    Wenbin Wang, Liang Ding, Minyan Zeng, Xiabin Zhou, Li Shen, Yong Luo, Wei Yu, and Dacheng Tao. Divide, conquer and combine: A training-free framework for high-resolution image perception in multimodal large language models. InProceedings of the AAAI Conference on Artificial Intelligence, pp. 7907–7915, 2025b. Lai Wei, Liangbo He, Jun Lan, Lingzhong Dong, ...

    work page arXiv

  20. [20]

    Qwen3 Technical Report

    An Yang et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans?arXiv preprint arXiv:2408.13257, 2024

    Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qingsong Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans?arXiv preprint arXiv:2408.13257,

    work page arXiv

  22. [22]

    Thyme: Think beyond images.arXiv preprint arXiv:2508.11630, 2025

    Yi-Fan Zhang, Xingyu Lu, Shukang Yin, Chaoyou Fu, Wei Chen, Xiao Hu, Bin Wen, Kaiyu Jiang, Changyi Liu, Tianke Zhang, et al. Thyme: Think beyond images.arXiv preprint arXiv:2508.11630, 2025a. Yifan Zhang, Liang Hu, Haofeng Sun, Peiyu Wang, Yichen Wei, Shukang Yin, Jiangbo Pei, Wei Shen, Peng Xia, Yi Peng, et al. Skywork-r1v4: Toward agentic multimodal int...

    work page arXiv

  23. [23]

    arXiv preprint arXiv:2408.05517 (2024),https://arxiv.org/abs/ 2408.05517

    URLhttps://arxiv.org/abs/2408.05517. Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai Xu, Le Yang, Chao Shen, and Xing Yu. Deepeyes: Incentivizing" thinking with images" via reinforcement learning.arXiv preprint arXiv:2505.14362,

    work page arXiv

  24. [24]

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479,

    work page internal anchor Pith review arXiv