VISOR: Agentic Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning

Dawei Yin; Jiulong Wu; Jizhou Huang; Lingyong Yan; Min Cao; Yucheng Shen

arxiv: 2604.09508 · v1 · submitted 2026-04-10 · 💻 cs.CV · cs.AI

VISOR: Agentic Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning

Yucheng Shen , Jiulong Wu , Jizhou Huang , Dawei Yin , Lingyong Yan , Min Cao This is my paper

Pith reviewed 2026-05-10 17:40 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords visual retrieval-augmented generationagentic visual reasoningevidence sparsitysearch driftstructured evidence spacedynamic trajectoryreinforcement learning for agentsvision-language models

0 comments

The pith

A single-agent framework uses structured evidence spaces and sliding-window trajectories to fix sparsity and drift in visual document reasoning.

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

The paper sets out to show that existing agentic systems for retrieving and reasoning over visual documents fail on complex queries because key evidence stays scattered across pages and agents lose their way as more visual information piles up. It introduces a unified framework that builds a progressive evidence space, corrects imprecise visual actions, and maintains a focused trajectory by sliding the context window while repeatedly injecting the original intent. If these changes work, agents could reliably chain together multi-page visual clues without overload or deviation, making vision-language models practical for detailed documents like reports or presentations. Readers should care because current approaches either need multiple agents or degrade quickly on long horizons, limiting real use cases.

Core claim

VISOR is a unified single-agent framework that features a structured Evidence Space for progressive cross-page reasoning, a Visual Action Evaluation and Correction mechanism to manage visual actions, and a Dynamic Trajectory with Sliding Window and Intent Injection to mitigate search drift by anchoring the evidence space while discarding earlier raw interactions and preventing context from being overwhelmed by visual tokens. The system is trained using a Group Relative Policy Optimization-based Reinforcement Learning pipeline with state masking and credit assignment tailored for dynamic context reconstruction.

What carries the argument

The Dynamic Trajectory with Sliding Window and Intent Injection, which anchors the evidence space and discards earlier raw interactions to keep the agent focused on its objective despite accumulating visual tokens.

If this is right

Progressive cross-page reasoning becomes feasible without processing each page in isolation.
Misuse of visual actions is reduced, preserving retrieval quality on fine-grained image details.
Context overload is avoided so the agent stays aligned with the original search goal over long sequences.
Reinforcement learning with state masking enables stable training on reconstructed dynamic contexts.
The single-agent design delivers higher performance and efficiency than prior multi-step visual agents on long-horizon tasks.

Where Pith is reading between the lines

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

The same sliding-window and intent-injection pattern could be tested in non-visual long-horizon agents to manage context without full history.
Single-agent designs with explicit memory reconstruction may prove more stable than multi-agent handoffs when evidence is spatially or temporally scattered.
The approach suggests that intent re-injection at each step could help other retrieval systems avoid gradual topic drift in extended interactions.

Load-bearing premise

That the evidence space, action correction, and sliding trajectory plus the tailored reinforcement learning will reduce sparsity and drift without introducing new failure modes or high extra cost.

What would settle it

An evaluation on queries whose required visual evidence spans more than ten distant pages in a single document, where the agent either retrieves incorrect pages or fails to chain the reasoning steps correctly.

Figures

Figures reproduced from arXiv: 2604.09508 by Dawei Yin, Jiulong Wu, Jizhou Huang, Lingyong Yan, Min Cao, Yucheng Shen.

**Figure 2.** Figure 2: Overview of VISOR. At each step 𝑖, the agent produces a ⟨think⟩. . .⟨action⟩ response. Evidence extracted from the reasoning trace is accumulated in a structured Evidence Collection space E. The action space comprises three operations: search, crop, and answer. The agent loop context is reconstructed at each turn: the user query and E are always pinned at the top, while only the last 𝑊 turns of raw interac… view at source ↗

**Figure 3.** Figure 3: Accuracy of Qwen2.5-VL-7B on SlideVQA (500 sam [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Breakdown of average per-sample inference latency [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: The prompt template used for LLM-as-Judge evaluation. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: A representative case illustrating redundant crop usage in VRAG-RL. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 7.** Figure 7: A success case of VISOR on a multi-hop SlideVQA question. Retrieved page images are shown at the top for layout [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

**Figure 8.** Figure 8: A failure case of VISOR on a multi-hop MMLongBench question. Both reference pages are retrieved correctly, but the [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

**Figure 9.** Figure 9: All prompt templates used in VISOR. Each template corresponds to a distinct interaction event in the agent loop. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

read the original abstract

Visual Retrieval-Augmented Generation (VRAG) empowers Vision-Language Models to retrieve and reason over visually rich documents. To tackle complex queries requiring multi-step reasoning, agentic VRAG systems interleave reasoning with iterative retrieval.. However, existing agentic VRAG faces two critical bottlenecks. (1) Visual Evidence Sparsity: key evidence is scattered across pages yet processed in isolation, hindering cross-page reasoning; moreover, fine-grained intra-image evidence often requires precise visual actions, whose misuse degrades retrieval quality; (2) Search Drift in Long Horizons: the accumulation of visual tokens across retrieved pages dilutes context and causes cognitive overload, leading agents to deviate from their search objective. To address these challenges, we propose VISOR (Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning), a unified single-agent framework. VISOR features a structured Evidence Space for progressive cross-page reasoning, coupled with a Visual Action Evaluation and Correction mechanism to manage visual actions. Additionally, we introduce a Dynamic Trajectory with Sliding Window and Intent Injection to mitigate search drift. They anchor the evidence space while discarding earlier raw interactions, preventing context from being overwhelmed by visual tokens. We train VISOR using a Group Relative Policy Optimization-based Reinforcement Learning (GRPO-based RL) pipeline with state masking and credit assignment tailored for dynamic context reconstruction. Extensive experiments on ViDoSeek, SlideVQA, and MMLongBench demonstrate that VISOR achieves state-of-the-art performance with superior efficiency for long-horizon visual reasoning tasks.

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.

Desk Editor's Note private letter to a colleague

VISOR adds a structured evidence space, action correction, and sliding-window trajectory to keep agentic visual RAG on track over long horizons, but the SOTA claims need the actual numbers to evaluate.

read the letter

VISOR targets two real issues in agentic visual retrieval-augmented generation: evidence scattered across pages that gets processed in isolation, and agents drifting off course as visual tokens pile up in long searches. The paper proposes a single-agent framework with a structured Evidence Space that builds cross-page reasoning progressively, a Visual Action Evaluation and Correction loop to fix bad visual actions, and a Dynamic Trajectory using a sliding window plus intent injection to drop old raw interactions while anchoring the search objective. They train the whole thing with GRPO-based RL that includes state masking and credit assignment for the changing context. These pieces form a new synthesis that directly addresses the bottlenecks mentioned in the abstract. The approach is straightforward and avoids over-engineering the agent loop, which is a plus for practical document-understanding work. The sliding window idea in particular looks like a clean way to manage context overload without losing the trajectory. The soft spot is the experimental side. The abstract asserts state-of-the-art results and efficiency gains on ViDoSeek, SlideVQA, and MMLongBench, yet the summary gives no quantitative deltas, baseline comparisons, or ablation breakdowns. If the full paper supplies those with clear error analysis and reproducible training details, the contribution lands; otherwise the gains could be modest or setup-specific. This is aimed at people building agentic multimodal systems for long visual documents. It deserves a serious referee because the mechanisms are logically consistent and target documented problems, even if the results require close checking for reproducibility and effect size. I would send it to peer review.

Referee Report

1 major / 0 minor

Summary. The paper proposes VISOR, a unified single-agent framework for agentic Visual Retrieval-Augmented Generation (VRAG) to address visual evidence sparsity and search drift in long-horizon visual reasoning. It introduces a structured Evidence Space for cross-page reasoning, Visual Action Evaluation and Correction for managing visual actions, a Dynamic Trajectory with Sliding Window and Intent Injection to prevent context overload, and GRPO-based RL training with state masking and credit assignment. The abstract claims state-of-the-art performance and superior efficiency on ViDoSeek, SlideVQA, and MMLongBench.

Significance. If the empirical results and ablations hold, the framework could meaningfully advance agentic VRAG by providing concrete mechanisms to handle scattered visual evidence and maintain objective focus over long trajectories, offering a more efficient alternative to multi-agent or heavily token-intensive approaches in visual document tasks.

major comments (1)

[Abstract] Abstract: the claim that VISOR 'achieves state-of-the-art performance with superior efficiency' is unsupported by any quantitative numbers, baseline comparisons, ablation studies, or error analysis. Without these, it is impossible to assess whether the proposed components (Evidence Space, action correction, sliding-window trajectory, and GRPO RL) actually mitigate the stated bottlenecks or introduce new failure modes.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thoughtful review and constructive criticism. We address the major comment point-by-point below and outline the revisions we will make to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that VISOR 'achieves state-of-the-art performance with superior efficiency' is unsupported by any quantitative numbers, baseline comparisons, ablation studies, or error analysis. Without these, it is impossible to assess whether the proposed components (Evidence Space, action correction, sliding-window trajectory, and GRPO RL) actually mitigate the stated bottlenecks or introduce new failure modes.

Authors: We agree that the abstract would be strengthened by including specific quantitative evidence to support the performance claims. The main body of the manuscript contains the requested elements: quantitative comparisons to baselines are presented in the experimental results on ViDoSeek, SlideVQA, and MMLongBench; ablation studies evaluate each proposed component (Evidence Space structuring, Visual Action Evaluation and Correction, Dynamic Trajectory with Sliding Window, and GRPO-based RL training); and discussions of how these address visual evidence sparsity and search drift, including potential limitations, are provided. To directly address the referee's concern, we will revise the abstract to incorporate key numerical results and efficiency metrics from our experiments. This will enable readers to immediately assess the effectiveness of the components without needing to refer to the full text initially. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces VISOR as a proposed single-agent framework consisting of explicitly described architectural components (structured Evidence Space, Visual Action Evaluation and Correction, Dynamic Trajectory with Sliding Window and Intent Injection) plus a GRPO-based RL training pipeline with state masking. These mechanisms are motivated by two stated bottlenecks and are evaluated via standard empirical benchmarks (ViDoSeek, SlideVQA, MMLongBench). No mathematical derivation, equations, or parameter-fitting steps are presented that reduce by construction to the claimed outputs. No load-bearing self-citations, uniqueness theorems, or ansatzes imported from prior author work appear in the abstract or described framework. The central claims therefore rest on independent architectural proposals and external benchmark results rather than self-referential definitions or fitted inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an applied systems paper whose contributions are architectural mechanisms and a training pipeline rather than mathematical derivations. No explicit free parameters, axioms, or independently evidenced invented entities are described in the abstract.

pith-pipeline@v0.9.0 · 5588 in / 1369 out tokens · 121883 ms · 2026-05-10T17:40:36.808841+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 14 internal anchors

[1]

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023
[2]

Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report.arXiv preprint arXiv:2309.16609(2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023
[3]

Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al . 2025. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[4]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. 2025. Qwen2.5-VL Technical Report.arXiv preprint arXiv:2502.13923(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[5]

Haizhou Du and Wenhao Li. 2026. M3RAG: Orchestrating Multi-agent Reason- ing for Multi-hop, Multi-modal Understanding. InInternational Conference on Multimedia Modeling. Springer, 364–378

work page 2026
[6]

Manuel Faysse, Hugues Sibille, Tony Wu, Bilel Omrani, Gautier Viaud, Céline Hudelot, and Pierre Colombo. 2024. Colpali: Efficient document retrieval with vision language models.arXiv preprint arXiv:2407.01449(2024)

work page internal anchor Pith review arXiv 2024
[7]

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, Haofen Wang, et al. 2023. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997 2, 1 (2023), 32

work page internal anchor Pith review Pith/arXiv arXiv 2023
[8]

Xinyu Geng, Peng Xia, Zhen Zhang, Xinyu Wang, Qiuchen Wang, Ruixue Ding, Chenxi Wang, Jialong Wu, Yida Zhao, Kuan Li, et al. 2025. Webwatcher: Breaking new frontier of vision-language deep research agent.arXiv preprint arXiv:2508.05748(2025)

work page internal anchor Pith review arXiv 2025
[9]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[10]

Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. 2022. Layoutlmv3: Pre-training for document ai with unified text and image masking. InProceedings of the 30th ACM international conference on multimedia. 4083–4091

work page 2022
[11]

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card.arXiv preprint arXiv:2412.16720(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[12]

Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanmin Wu, Jiayi Lei, Pengshuo Qiu, Pan Lu, Zehui Chen, Chaoyou Fu, Guanglu Song, et al. 2024. Mmsearch: Bench- marking the potential of large models as multi-modal search engines.arXiv preprint arXiv:2409.12959(2024)

work page arXiv 2024
[13]

Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516(2025)

work page Pith review arXiv 2025
[14]

Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. 2022. Ocr-free document understanding transformer. InEuropean Confer- ence on Computer Vision. Springer, 498–517

work page 2022
[15]

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems33 (2020), 9459–9474

work page 2020
[16]

Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[17]

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning.Advances in neural information processing systems36 (2023), 34892–34916

work page 2023
[18]

Keliang Liu, Zizhi Chen, Mingcheng Li, Jingqun Tang, Dingkang Yang, and Lihua Zhang. 2025. Resolving evidence sparsity: Agentic context engineering for long-document understanding.arXiv preprint arXiv:2511.22850(2025)

work page arXiv 2025
[19]

Ziyu Liu, Yuhang Zang, Yushan Zou, Zijian Liang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. 2025. Visual agentic reinforcement fine-tuning.arXiv preprint arXiv:2505.14246(2025)

work page arXiv 2025
[20]

Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li, Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, et al. 2024. Mmlongbench-doc: Benchmarking long-context document understanding with visualizations.Advances in Neural Information Processing Systems37 (2024), 95963–96010

work page 2024
[21]

Chunyi Peng, Zhipeng Xu, Zhenghao Liu, Yishan Li, Yukun Yan, Shuo Wang, Zhiyuan Liu, Yu Gu, Minghe Yu, Ge Yu, et al. 2025. Learning to route queries across knowledge bases for step-wise retrieval-augmented reasoning.arXiv preprint arXiv:2505.22095(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[22]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[23]

Yubo Sun, Chunyi Peng, Yukun Yan, Shi Yu, Zhenghao Liu, Chi Chen, Zhiyuan Liu, and Maosong Sun. 2025. VisRAG 2.0: Evidence-Guided Multi-Image Reasoning in Visual Retrieval-Augmented Generation.arXiv preprint arXiv:2510.09733(2025)

work page arXiv 2025
[24]

Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. 2023. Slidevqa: A dataset for document visual question an- swering on multiple images. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 13636–13645

work page 2023
[25]

Qiuchen Wang, Ruixue Ding, Zehui Chen, Weiqi Wu, Shihang Wang, Pengjun Xie, and Feng Zhao. 2025. Vidorag: Visual document retrieval-augmented generation via dynamic iterative reasoning agents. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 9124–9145

work page 2025
[26]

Qiuchen Wang, Ruixue Ding, Yu Zeng, Zehui Chen, Lin Chen, Shihang Wang, Pengjun Xie, Fei Huang, and Feng Zhao. 2025. Vrag-rl: Empower vision- perception-based rag for visually rich information understanding via iterative reasoning with reinforcement learning.arXiv preprint arXiv:2505.22019(2025)

work page arXiv 2025
[27]

Jinming Wu, Zihao Deng, Wei Li, Yiding Liu, Bo You, Bo Li, Zejun Ma, and Ziwei Liu. 2025. Mmsearch-r1: Incentivizing lmms to search.arXiv preprint arXiv:2506.20670(2025)

work page internal anchor Pith review arXiv 2025
[28]

Jiulong Wu, Yucheng Shen, Lingyong Yan, Haixin Sun, Deguo Xia, Jizhou Huang, and Min Cao. 2025. Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis.arXiv preprint arXiv:2511.10254(2025)

work page arXiv 2025
[29]

Xiao Yang, Kai Sun, Hao Xin, Yushi Sun, Nikita Bhalla, Xiangsen Chen, Sa- jal Choudhary, Rongze D Gui, Ziran W Jiang, Ziyu Jiang, et al . 2024. Crag- comprehensive rag benchmark.Advances in Neural Information Processing Sys- tems37 (2024), 10470–10490

work page 2024
[30]

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations

work page 2022
[31]

VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents

Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zheng- hao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, et al . 2024. Visrag: Vision-based retrieval-augmented generation on multi-modality documents.arXiv preprint arXiv:2410.10594(2024). A Search Engine and Crop-and-Zoom Tool Search Engine.We use ColQwen2.5-v0.1 [ 6] as our retrieval back- bone. ...

work page internal anchor Pith review arXiv 2024
[32]

Every response must start with <think> </think> where you reason about what you see and what to do next

work page
[33]

Each search returns one new image; if you repeat a query, you will get a different image from the same document

After thinking, output exactly one action: - <search>query</search> to retrieve images. Each search returns one new image; if you repeat a query, you will get a different image from the same document. Use the original question as your query unless you have a specific reason to change it. - <bbox>[x1,y1,x2,y2]</bbox> to zoom into an unclear region (normali...

work page
[34]

Before answering, you must do one final search using the original question to verify your answer. After receiving the new image, give your <answer> immediately — unless the new image provides a directly conflicting answer to the question, in which case search once more and then give your <answer> immediately regardless

work page
[35]

This image does not contain information related to the question

When given an image, analyze it fully in <think> </think> and extract every potentially useful piece of information — your thoughts will be recorded into a COLLECTED EVIDENCE table for later reference, so be as thorough as possible. If the image contains no relevant information, explicitly state that (e.g., "This image does not contain information related...

work page 2007

[1] [1]

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023

[2] [2]

Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report.arXiv preprint arXiv:2309.16609(2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023

[3] [3]

Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al . 2025. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025

[4] [4]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. 2025. Qwen2.5-VL Technical Report.arXiv preprint arXiv:2502.13923(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025

[5] [5]

Haizhou Du and Wenhao Li. 2026. M3RAG: Orchestrating Multi-agent Reason- ing for Multi-hop, Multi-modal Understanding. InInternational Conference on Multimedia Modeling. Springer, 364–378

work page 2026

[6] [6]

Manuel Faysse, Hugues Sibille, Tony Wu, Bilel Omrani, Gautier Viaud, Céline Hudelot, and Pierre Colombo. 2024. Colpali: Efficient document retrieval with vision language models.arXiv preprint arXiv:2407.01449(2024)

work page internal anchor Pith review arXiv 2024

[7] [7]

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, Haofen Wang, et al. 2023. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997 2, 1 (2023), 32

work page internal anchor Pith review Pith/arXiv arXiv 2023

[8] [8]

Xinyu Geng, Peng Xia, Zhen Zhang, Xinyu Wang, Qiuchen Wang, Ruixue Ding, Chenxi Wang, Jialong Wu, Yida Zhao, Kuan Li, et al. 2025. Webwatcher: Breaking new frontier of vision-language deep research agent.arXiv preprint arXiv:2508.05748(2025)

work page internal anchor Pith review arXiv 2025

[9] [9]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025

[10] [10]

Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. 2022. Layoutlmv3: Pre-training for document ai with unified text and image masking. InProceedings of the 30th ACM international conference on multimedia. 4083–4091

work page 2022

[11] [11]

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card.arXiv preprint arXiv:2412.16720(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024

[12] [12]

Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanmin Wu, Jiayi Lei, Pengshuo Qiu, Pan Lu, Zehui Chen, Chaoyou Fu, Guanglu Song, et al. 2024. Mmsearch: Bench- marking the potential of large models as multi-modal search engines.arXiv preprint arXiv:2409.12959(2024)

work page arXiv 2024

[13] [13]

Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516(2025)

work page Pith review arXiv 2025

[14] [14]

Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. 2022. Ocr-free document understanding transformer. InEuropean Confer- ence on Computer Vision. Springer, 498–517

work page 2022

[15] [15]

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems33 (2020), 9459–9474

work page 2020

[16] [16]

Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024

[17] [17]

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning.Advances in neural information processing systems36 (2023), 34892–34916

work page 2023

[18] [18]

Keliang Liu, Zizhi Chen, Mingcheng Li, Jingqun Tang, Dingkang Yang, and Lihua Zhang. 2025. Resolving evidence sparsity: Agentic context engineering for long-document understanding.arXiv preprint arXiv:2511.22850(2025)

work page arXiv 2025

[19] [19]

Ziyu Liu, Yuhang Zang, Yushan Zou, Zijian Liang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. 2025. Visual agentic reinforcement fine-tuning.arXiv preprint arXiv:2505.14246(2025)

work page arXiv 2025

[20] [20]

Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li, Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, et al. 2024. Mmlongbench-doc: Benchmarking long-context document understanding with visualizations.Advances in Neural Information Processing Systems37 (2024), 95963–96010

work page 2024

[21] [21]

Chunyi Peng, Zhipeng Xu, Zhenghao Liu, Yishan Li, Yukun Yan, Shuo Wang, Zhiyuan Liu, Yu Gu, Minghe Yu, Ge Yu, et al. 2025. Learning to route queries across knowledge bases for step-wise retrieval-augmented reasoning.arXiv preprint arXiv:2505.22095(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025

[22] [22]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024

[23] [23]

Yubo Sun, Chunyi Peng, Yukun Yan, Shi Yu, Zhenghao Liu, Chi Chen, Zhiyuan Liu, and Maosong Sun. 2025. VisRAG 2.0: Evidence-Guided Multi-Image Reasoning in Visual Retrieval-Augmented Generation.arXiv preprint arXiv:2510.09733(2025)

work page arXiv 2025

[24] [24]

Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. 2023. Slidevqa: A dataset for document visual question an- swering on multiple images. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 13636–13645

work page 2023

[25] [25]

Qiuchen Wang, Ruixue Ding, Zehui Chen, Weiqi Wu, Shihang Wang, Pengjun Xie, and Feng Zhao. 2025. Vidorag: Visual document retrieval-augmented generation via dynamic iterative reasoning agents. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 9124–9145

work page 2025

[26] [26]

Qiuchen Wang, Ruixue Ding, Yu Zeng, Zehui Chen, Lin Chen, Shihang Wang, Pengjun Xie, Fei Huang, and Feng Zhao. 2025. Vrag-rl: Empower vision- perception-based rag for visually rich information understanding via iterative reasoning with reinforcement learning.arXiv preprint arXiv:2505.22019(2025)

work page arXiv 2025

[27] [27]

Jinming Wu, Zihao Deng, Wei Li, Yiding Liu, Bo You, Bo Li, Zejun Ma, and Ziwei Liu. 2025. Mmsearch-r1: Incentivizing lmms to search.arXiv preprint arXiv:2506.20670(2025)

work page internal anchor Pith review arXiv 2025

[28] [28]

Jiulong Wu, Yucheng Shen, Lingyong Yan, Haixin Sun, Deguo Xia, Jizhou Huang, and Min Cao. 2025. Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis.arXiv preprint arXiv:2511.10254(2025)

work page arXiv 2025

[29] [29]

Xiao Yang, Kai Sun, Hao Xin, Yushi Sun, Nikita Bhalla, Xiangsen Chen, Sa- jal Choudhary, Rongze D Gui, Ziran W Jiang, Ziyu Jiang, et al . 2024. Crag- comprehensive rag benchmark.Advances in Neural Information Processing Sys- tems37 (2024), 10470–10490

work page 2024

[30] [30]

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations

work page 2022

[31] [31]

VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents

Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zheng- hao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, et al . 2024. Visrag: Vision-based retrieval-augmented generation on multi-modality documents.arXiv preprint arXiv:2410.10594(2024). A Search Engine and Crop-and-Zoom Tool Search Engine.We use ColQwen2.5-v0.1 [ 6] as our retrieval back- bone. ...

work page internal anchor Pith review arXiv 2024

[32] [32]

Every response must start with <think> </think> where you reason about what you see and what to do next

work page

[33] [33]

Each search returns one new image; if you repeat a query, you will get a different image from the same document

After thinking, output exactly one action: - <search>query</search> to retrieve images. Each search returns one new image; if you repeat a query, you will get a different image from the same document. Use the original question as your query unless you have a specific reason to change it. - <bbox>[x1,y1,x2,y2]</bbox> to zoom into an unclear region (normali...

work page

[34] [34]

Before answering, you must do one final search using the original question to verify your answer. After receiving the new image, give your <answer> immediately — unless the new image provides a directly conflicting answer to the question, in which case search once more and then give your <answer> immediately regardless

work page

[35] [35]

This image does not contain information related to the question

When given an image, analyze it fully in <think> </think> and extract every potentially useful piece of information — your thoughts will be recorded into a COLLECTED EVIDENCE table for later reference, so be as thorough as possible. If the image contains no relevant information, explicitly state that (e.g., "This image does not contain information related...

work page 2007