pith. machine review for the scientific record. sign in

arxiv: 2605.04304 · v1 · submitted 2026-05-05 · 💻 cs.CV · cs.CL

Recognition: unknown

Hierarchical Visual Agent: Managing Contexts in Joint Image-Text Space for Advanced Chart Reasoning

Qihua Dong , Ruozhen He , Junwen Chen , Yizhou Wang , Xu Ma , Songyao Jiang , Yun Fu
Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:03 UTC · model grok-4.3

classification 💻 cs.CV cs.CL
keywords chart reasoninghierarchical agentsmultimodal reasoningvisual context managementquestion answeringcontext distillation
0
0 comments X

The pith

A hierarchical agent maintains compact joint image-text contexts to improve multi-step reasoning on complex charts.

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

The paper proposes a framework where a high-level manager creates plans and keeps only essential information in separate visual and textual contexts, while worker agents execute steps and can zoom into specific image regions. This structure targets the gap in existing multimodal models that handle single plots well but falter on sequential reasoning across multiple subplots. If correct, the approach would let agents process detailed charts more reliably by avoiding overload from full images and by distilling context at each step. Readers should care because many real-world data analysis tasks involve exactly these multi-element visuals and chained questions.

Core claim

The central claim is that iteratively building and updating a working context in joint image-text space, with a manager maintaining a compact distilled record and workers using a zoom-in tool to restrict visual input, enables stronger performance on advanced chart question answering than flat multimodal large language models.

What carries the argument

The manager-worker hierarchy that keeps separate visual and textual contexts and applies a zoom-in tool to scope visual attention to relevant chart elements.

If this is right

  • Hierarchical planning plus scoped visual context together produce larger gains than either alone on multi-plot reasoning tasks.
  • Distilling context at each iteration prevents overload while preserving information needed for later steps.
  • The three components—architecture, visual scoping, and context distillation—each add measurable independent value according to the reported ablations.
  • The method yields consistent accuracy lifts over strong multimodal baselines on the CharXiv reasoning subset.

Where Pith is reading between the lines

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

  • The same context-management pattern could be applied to other sequential visual tasks such as document navigation or diagram-based problem solving.
  • Explicit separation of planning and perception might reduce error accumulation in longer reasoning chains beyond charts.
  • Testing the framework on charts with increasing numbers of subplots would reveal whether the compactness benefit scales.

Load-bearing premise

The manager can always create correct plans and retain every necessary detail in its compact context without loss, and the zoom-in tool can isolate exactly the visual parts required for each worker step.

What would settle it

An ablation on the CharXiv reasoning subset that removes the hierarchy or the zoom-in tool and measures whether performance drops to baseline levels, or a test on new multi-subplot charts where the agent loses critical details across reasoning steps.

Figures

Figures reproduced from arXiv: 2605.04304 by Junwen Chen, Qihua Dong, Ruozhen He, Songyao Jiang, Xu Ma, Yizhou Wang, Yun Fu.

Figure 1
Figure 1. Figure 1: Comparison of chart reasoning paradigms. (a) Chain-of-Thought (CoT) reasons sequentially over a global image but is often distracted by irrelevant vi￾sual content. (b) Thinking with Images (Interleaved) iteratively acquires visual crops but appends them to a monotonically growing context, leading to clutter and distraction. (c) HIERVA (Ours) uses hierarchical visual agents that maintain a joint image-text … view at source ↗
Figure 2
Figure 2. Figure 2: Case study comparing reasoning approaches. We illustrate how different methods handle an advanced chart reasoning question. CoT reasons sequentially over the global image but struggles with fine-grained details. Thinking w/ Images iteratively zooms into regions but accumulates all crops and intermediate reasoning in a growing context, leading to errors. HIERVA manages the context effectively by using hiera… view at source ↗
read the original abstract

Advanced chart question answering requires both precise perception of small visual elements and multi-step reasoning across several subplots. While existing MLLMs are strong at understanding single plots, they often struggle with multi-step reasoning across multiple subplots. We propose HierVA, a hierarchical visual agent framework for chart reasoning that iteratively constructs and updates a working context in a joint image--text space. A high-level manager generates plans and maintains a compact context containing only key information, while specialized workers perform reasoning, gather evidence, and return results. In particular, the agent maintains separate visual and textual contexts, using a zoom-in tool to restrict the visual context. Experiments on the CharXiv reasoning subset demonstrate consistent improvements over strong multimodal baselines, and ablation studies verify that hierarchical architecture, scoped visual context, and distilled context contribute complementary gains.

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 / 0 minor

Summary. The manuscript introduces HierVA, a hierarchical visual agent framework for advanced chart reasoning. A high-level manager generates plans and maintains a compact working context in joint image-text space, while specialized workers execute reasoning, gather evidence, and return results; a zoom-in tool is used to scope the visual context. Experiments on the CharXiv reasoning subset are claimed to demonstrate consistent improvements over strong multimodal baselines, with ablation studies verifying complementary gains from the hierarchical architecture, scoped visual context, and distilled context.

Significance. If the reported results hold with detailed validation, the hierarchical separation of planning from execution and the joint management of scoped visual and distilled textual contexts could offer a practical advance for multimodal models on complex chart tasks that require both fine-grained perception of small elements and multi-step reasoning across subplots. This design directly targets limitations of current MLLMs and, if the ablations confirm non-redundant contributions, would provide a reusable template for context-efficient agentic systems in vision-language reasoning.

major comments (2)
  1. [Experiments] Experiments section: the central claim of 'consistent improvements over strong multimodal baselines' and 'complementary gains' from the three ablated components rests entirely on experimental outcomes, yet the manuscript text supplies no quantitative metrics, baseline names or scores, dataset statistics, statistical significance tests, or error analysis. Without these, the magnitude, reliability, and reproducibility of the gains cannot be assessed and the claim remains unverifiable.
  2. [Method] Method section (description of manager and zoom-in tool): the framework presupposes that the high-level manager reliably produces plans that preserve all necessary information in a compact context and that the zoom-in tool isolates relevant visual elements without introducing errors or omissions. No robustness analysis, failure-case discussion, or empirical check of information retention is provided, which is load-bearing for the asserted benefits of the hierarchical design and scoped contexts.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below and describe the revisions we will incorporate.

read point-by-point responses

  1. Referee: Experiments section: the central claim of 'consistent improvements over strong multimodal baselines' and 'complementary gains' from the three ablated components rests entirely on experimental outcomes, yet the manuscript text supplies no quantitative metrics, baseline names or scores, dataset statistics, statistical significance tests, or error analysis. Without these, the magnitude, reliability, and reproducibility of the gains cannot be assessed and the claim remains unverifiable.

    Authors: We agree that the main-text narrative does not explicitly enumerate the numerical results, baseline identities, dataset statistics, significance tests, or error analysis, even though these appear in the accompanying tables and figures. In the revised manuscript we will expand the Experiments section to state the CharXiv reasoning-subset statistics, list all baseline models and their exact scores, report the observed improvements with numerical deltas, include any statistical significance results, and add a concise error analysis of recurring failure modes. These additions will make the central claims directly verifiable from the text. revision: yes

  2. Referee: Method section (description of manager and zoom-in tool): the framework presupposes that the high-level manager reliably produces plans that preserve all necessary information in a compact context and that the zoom-in tool isolates relevant visual elements without introducing errors or omissions. No robustness analysis, failure-case discussion, or empirical check of information retention is provided, which is load-bearing for the asserted benefits of the hierarchical design and scoped contexts.

    Authors: We acknowledge the absence of a dedicated robustness analysis, failure-case discussion, or quantitative check of information retention for the manager and zoom-in tool. In the revision we will add a short subsection (or paragraph within the Method/Experiments) that discusses potential failure modes—such as omitted plan elements or incomplete visual scoping—supported by qualitative examples drawn from our development set. We will also report an empirical information-retention check on a held-out sample, using either human annotation or a proxy metric that compares the distilled context against ground-truth key facts. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical validation

full rationale

The paper describes an architectural framework (HierVA) for chart reasoning via a hierarchical manager-worker agent that maintains scoped visual and textual contexts, then validates it through experiments on CharXiv and ablation studies showing complementary gains. No derivation chain, equations, fitted parameters, or first-principles results are present that could reduce to self-referential inputs by construction. Claims of improvement are grounded in reported experimental outcomes rather than any self-definition, renamed known results, or load-bearing self-citations. The central premise is externally falsifiable via the ablations and baselines, making the work self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim depends on the unproven effectiveness of the newly introduced hierarchical decomposition and context-scoping mechanisms; no free parameters are mentioned, but the design rests on domain assumptions about MLLM limitations and the utility of agentic iteration.

axioms (1)
  • domain assumption Existing MLLMs struggle with multi-step reasoning across multiple subplots
    Explicitly stated as motivation in the abstract for needing the hierarchical agent.
invented entities (2)
  • HierVA hierarchical visual agent no independent evidence
    purpose: Iteratively constructs and updates working context in joint image-text space
    Core new framework proposed to address chart reasoning limitations.
  • Zoom-in tool no independent evidence
    purpose: Restricts the visual context to relevant elements
    Specialized mechanism introduced as part of the agent architecture.

pith-pipeline@v0.9.0 · 5453 in / 1372 out tokens · 110836 ms · 2026-05-08T17:03:28.645809+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

40 extracted references · 29 canonical work pages · 13 internal anchors

  1. [7]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. arXiv preprint arXiv:2201.11903 , year =

    work page internal anchor Pith review arXiv

  2. [8]

    2025 , month = apr, url =

    OpenAI o3 and o4-mini System Card , author =. 2025 , month = apr, url =

    2025

  3. [9]

    DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning

    DeepEyes: Incentivizing ``Thinking with Images'' via Reinforcement Learning , author =. arXiv preprint arXiv:2505.14362 , year =

    work page internal anchor Pith review arXiv

  4. [10]

    Mini-o3: Scaling up reasoning pat- terns and interaction turns for visual search.arXiv preprint arXiv:2509.07969, 2025

    Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search , author =. arXiv preprint arXiv:2509.07969 , year =

    work page arXiv

  5. [11]

    International Conference on Learning Representations (ICLR) , year =

    ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning , author =. International Conference on Learning Representations (ICLR) , year =

  6. [12]

    2025 , month = aug, day =

    OpenAI , title =. 2025 , month = aug, day =

    2025

  7. [16]

    Findings of the Association for Computational Linguistics: ACL 2024 , month = aug, year =

    ChartAssistant: A Universal Chart Multimodal Language Model via Chart-to-Table Pre-training and Multitask Instruction Tuning , author =. Findings of the Association for Computational Linguistics: ACL 2024 , month = aug, year =

    2024

  8. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Gupta, Tanmay and Kembhavi, Aniruddha , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2023 , pages =

    2023

  9. [25]

    L La V A- N e X T: Improved reasoning, OCR, and world knowledge

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Li, Bo and Zhang, Yuanhan and Shen, Sheng and Lee, Yong Jae. L La V A- N e X T: Improved reasoning, OCR, and world knowledge. 2024

    2024

  10. [32]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, and 2 others. 2025. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631

    work page internal anchor Pith review arXiv 2025

  11. [33]

    Lei Chen, Xuanle Zhao, Zhixiong Zeng, Jing Huang, Yufeng Zhong, and Lin Ma. 2025. Chart-r1: Chain-of-thought supervision and reinforcement for advanced chart reasoner. arXiv preprint arXiv:2507.15509

    work page arXiv 2025

  12. [34]

    Tanmay Gupta and Aniruddha Kembhavi. 2023. https://openaccess.thecvf.com/content/CVPR2023/html/Gupta_Visual_Programming_Compositional_Visual_Reasoning_Without_Training_CVPR_2023_paper.html Visual programming: Compositional visual reasoning without training . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages...

    2023

  13. [35]

    Mengkang Hu, Tianxing Chen, Qiguang Chen, Yao Mu, Wenqi Shao, and Ping Luo. 2024. https://doi.org/10.48550/arXiv.2408.09559 Hiagent: Hierarchical working memory management for solving long-horizon agent tasks with large language model . arXiv preprint arXiv:2408.09559

    work page doi:10.48550/arxiv.2408.09559 2024

  14. [36]

    Amanpreet Kaur, Ahmed Masry, Enamul Hoque, and Shafiq Joty. 2025. https://doi.org/10.48550/arXiv.2501.09007 Chartagent: A multimodal agent for visually grounded reasoning in complex chart question answering . arXiv preprint arXiv:2501.09007

    work page doi:10.48550/arxiv.2501.09007 2025

  15. [37]

    Xin Lai, Junyi Li, Wei Li, Tao Liu, Tianjian Li, and Hengshuang Zhao. 2025. https://doi.org/10.48550/arXiv.2509.07969 Mini-o3: Scaling up reasoning patterns and interaction turns for visual search . arXiv preprint arXiv:2509.07969

    work page doi:10.48550/arxiv.2509.07969 2025

  16. [38]

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. https://doi.org/10.48550/arXiv.2303.17760 Camel: Communicative agents for 'mind' exploration of large language model society . arXiv preprint arXiv:2303.17760. Accepted at NeurIPS 2023

    work page internal anchor Pith review doi:10.48550/arxiv.2303.17760 2023

  17. [39]

    Fangyu Liu, Julian Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. 2023 a . https://doi.org/10.18653/v1/2023.findings-acl.660 D e P lot: One-shot visual language reasoning by plot-to-table translation . In Findings of the Association for Computational Linguistics: ACL 2...

    work page doi:10.18653/v1/2023.findings-acl.660 2023

  18. [40]

    Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Eisenschlos. 2023 b . https://doi.org/10.18653/v1/2023.acl-long.714 M at C ha: Enhancing visual language pretraining with math reasoning and chart derendering . In Proceedings of the 61st Annual Meeting of the Association for Co...

    work page doi:10.18653/v1/2023.acl-long.714 2023

  19. [41]

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024. https://llava-vl.github.io/blog/2024-01-30-llava-next/ L la V a- N e X t: Improved reasoning, ocr, and world knowledge . LLaVA Blog

    2024

  20. [42]

    Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.906 U ni C hart: A universal vision-language pretrained model for chart comprehension and reasoning . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14662--14684, Singapore. Associa...

    work page doi:10.18653/v1/2023.emnlp-main.906 2023

  21. [43]

    Ahmed Masry, Do Long, Jia Tan, Shafiq Joty, and Enamul Hoque. 2022. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In International Conference on Learning Representations (ICLR)

    2022

  22. [44]

    Ahmed Masry, Mehrad Shahmohammadi, Md Rizwan Parvez, Enamul Hoque, and Shafiq Joty. 2024 a . https://doi.org/10.18653/v1/2024.findings-acl.619 C hart I nstruct: Instruction tuning for chart comprehension and reasoning . In Findings of the Association for Computational Linguistics: ACL 2024, pages 10387--10409, Bangkok, Thailand. Association for Computatio...

    work page doi:10.18653/v1/2024.findings-acl.619 2024

  23. [45]

    Ahmed Masry, Megh Thakkar, Aayush Bajaj, Aaryaman Kartha, Enamul Hoque, and Shafiq Joty. 2024 b . https://doi.org/10.48550/arXiv.2407.04172 Chartgemma: Visual instruction-tuning for chart reasoning in the wild . arXiv preprint arXiv:2407.04172

    work page doi:10.48550/arxiv.2407.04172 2024

  24. [46]

    Fanqing Meng, Wenqi Shao, Quanfeng Lu, Peng Gao, Kaipeng Zhang, Yu Qiao, and Ping Luo. 2024. https://doi.org/10.18653/v1/2024.findings-acl.463 Chartassistant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning . In Findings of the Association for Computational Linguistics: ACL 2024, pages 7775--780...

    work page doi:10.18653/v1/2024.findings-acl.463 2024

  25. [47]

    OpenAI. 2025 a . GPT-5 system card. https://openai.com/index/gpt-5-system-card/. Accessed: 2026-01-05

    2025

  26. [48]

    OpenAI. 2025 b . https://cdn.openai.com/pdf/2221c875-02dc-4789-800b-e7758f3722c1/o3-and-o4-mini-system-card.pdf Openai o3 and o4-mini system card . Technical report, OpenAI

    2025

  27. [49]

    MemGPT: Towards LLMs as Operating Systems

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. 2023. https://doi.org/10.48550/arXiv.2310.08560 Memgpt: Towards llms as operating systems . arXiv preprint arXiv:2310.08560

    work page internal anchor Pith review doi:10.48550/arxiv.2310.08560 2023

  28. [50]

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. arXiv preprint arXiv:2303.11366

    work page internal anchor Pith review arXiv 2023

  29. [51]

    Didac Suris, Sachit Menon, and Carl Vondrick. 2023. https://doi.org/10.48550/arXiv.2303.08128 Vipergpt: Visual inference via python execution for reasoning . arXiv preprint arXiv:2303.08128

    work page doi:10.48550/arxiv.2303.08128 2023

  30. [52]

    Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, and Ee-Peng Lim. 2023. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. arXiv preprint arXiv:2305.04091

    work page arXiv 2023

  31. [53]

    Zihan Wang, Ahmed Masry, Enamul Hoque, and Shafiq Joty. 2024. https://doi.org/10.48550/arXiv.2406.18521 Charxiv: Charting gaps in realistic chart understanding in multimodal llms . arXiv preprint arXiv:2406.18521

    work page doi:10.48550/arxiv.2406.18521 2024

  32. [54]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. https://doi.org/10.48550/arXiv.2201.11903 Chain-of-thought prompting elicits reasoning in large language models . arXiv preprint arXiv:2201.11903

    work page internal anchor Pith review doi:10.48550/arxiv.2201.11903 2022

  33. [55]

    Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. 2023 a . https://doi.org/10.48550/arXiv.2303.04671 Visual chatgpt: Talking, drawing and editing with visual foundation models . arXiv preprint arXiv:2303.04671

    work page internal anchor Pith review doi:10.48550/arxiv.2303.04671 2023

  34. [56]

    AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. 2023 b . https://doi.org/10.48550/arXiv.2308.08155 Autogen: Enabling next-gen llm applications via multi-agent conversation . arXiv preprint arXiv:2308.08155

    work page Pith review doi:10.48550/arxiv.2308.08155 2023

  35. [57]

    Binfeng Xu, Zhiyuan Peng, Bowen Lei, Subhabrata Mukherjee, Yuchen Liu, and Dongkuan Xu. 2023. https://doi.org/10.48550/arXiv.2305.18323 Rewoo: Decoupling reasoning from observations for efficient augmented language models . arXiv preprint arXiv:2305.18323

    work page internal anchor Pith review doi:10.48550/arxiv.2305.18323 2023

  36. [58]

    Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. 2023. https://doi.org/10.48550/arXiv.2303.11381 Mm-react: Prompting chatgpt for multimodal reasoning and action . arXiv preprint arXiv:2303.11381

    work page internal anchor Pith review doi:10.48550/arxiv.2303.11381 2023

  37. [59]

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. https://doi.org/10.48550/arXiv.2305.10601 Tree of thoughts: Deliberate problem solving with large language models . arXiv preprint arXiv:2305.10601. NeurIPS 2023 camera-ready version

    work page internal anchor Pith review doi:10.48550/arxiv.2305.10601 2023

  38. [60]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629

    work page internal anchor Pith review arXiv 2022

  39. [61]

    Recursive Language Models

    Alex L. Zhang, Tim Kraska, and Omar Khattab. 2025. Recursive language models. arXiv preprint arXiv:2512.24601

    work page Pith review arXiv 2025

  40. [62]

    Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai Xu, Le Yang, Chao Shen, and Xing Yu. 2025. https://doi.org/10.48550/arXiv.2505.14362 Deepeyes: Incentivizing ``thinking with images'' via reinforcement learning . arXiv preprint arXiv:2505.14362

    work page internal anchor Pith review doi:10.48550/arxiv.2505.14362 2025