Pith. sign in

REVIEW 4 major objections 5 minor 66 references

The paper claims a video deep-research agent can beat proprietary models by forcing visual grounding before web retrieval, and backs the claim with a leakage-filtered benchmark.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-05 04:38 UTC pith:KC5WV6PW

load-bearing objection A genuinely new task framing and a sensible training recipe, but the headline SOTA gap rests on a benchmark whose size, judge, and leakage controls are not yet credible. the 4 major comments →

arxiv 2608.03979 v1 pith:KC5WV6PW submitted 2026-08-04 cs.CV cs.AI

Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent

Zhen Fang , Yu Zeng , Wenxuan Huang , Yiming Zhao , Shiting Huang , Tianfei Ren , Qi Lu , Qingnan Ren
show 12 more authors
Qisheng Su Lionel Z. Wang Qingyu Yin Shuang Chen Zehui Chen Lin Chen Zhenfei Yin Yao Hu Shaohui Lin Wanli Ouyang Shaosheng Cao Feng Zhao
This is my paper
classification cs.CV cs.AI
keywords video deep researchmultimodal agentsvisual groundingtool-augmented reasoningparametric knowledge leakagegroup relative policy optimizationmulti-hop video QAopen-web exploration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that multimodal deep research can move from static images to continuous video, and that the right training pipeline—not raw model size—is what unlocks it. It argues that current agents fail this task for two reasons: they avoid visual tools in favor of text search (modality bias), and they answer from memorized knowledge instead of actually grounding in the video (parametric knowledge leakage). To counter both, the paper builds Video-DeepResearch: a decoupled perception-then-exploration pipeline that forces the agent to select keyframes and crop-search visual entities before any web text retrieval, trained with supervised fine-tuning followed by GRPO reinforcement learning. On a new human-AI collaborative benchmark of 200 multi-hop video questions that the paper claims provably require both visual search and external knowledge, the 35B variant reaches 64.0% average accuracy, beating Claude-4.5-Sonnet (59.0%), Gemini 2.5 Pro (57.5%), and GPT-5 (52.5%). If these results hold, they would show that agentic tool use for video can be taught at compact scale, and that existing benchmarks were partly measuring memorization rather than real video understanding.

Core claim

The central discovery is that video deep research can be operationalized as sequential grounding: first find the informative frames, then crop the salient entities and run visual search on them, and only then turn to text web search. The paper identifies two failures in current agents—visual-tool aversion and parametric-knowledge leakage—and shows that a decoupled perception-exploration pipeline with stage-wise tool unlocking, plus SFT and GRPO training, corrects both. On the paper's Video-DR-Bench, Video-DeepResearch-35B-A3B reaches 64.0% average accuracy, surpassing Claude-4.5-Sonnet (59.0%) by 5.0 points and clearly outperforming GPT-5 (52.5%) and Gemini 2.5 Pro (57.5%); the 30B-A3B varia

What carries the argument

The mechanism that carries the argument is the decoupled perception-exploration pipeline with stage-wise tool unlocking. The agent is first restricted to two visual tools—Select_Keyframe, which isolates an informative frame from the video stream, and Crop_Search, which crops a bounding box around a salient entity and runs a reverse-image web search—and is forced to exhaustively ground visual entities across frames before the action space expands to Text_Search and Visit. This ordering compels the agent to build concrete visual evidence before web retrieval. The two-stage training recipe then locks in the behavior: supervised fine-tuning on 7K curated trajectories plus 7K text-only QA instanc

Load-bearing premise

The central claim stands on the assumption that no Video-DR-Bench question can be answered from the model's memorized knowledge—the paper filters with four tool-free rollouts plus human verification, and the whole 'provably requires tools' guarantee depends on that filter being airtight.

What would settle it

First, resolve the printed inconsistency in benchmark size (200 in the abstract and conclusion, 100 in Section 3.4). Then take the actual benchmark, disable all tools, and let a strong model answer each question many times with varied decoding; if any non-trivial fraction is correct without tools, the 'provably requires both visual search and external knowledge' claim is falsified. A companion check is to search the benchmark videos' frames and source URLs in the SFT/GRPO training corpora for contamination.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Compact open models trained this way can outperform much larger proprietary models on this benchmark, so model scale alone is not enough for video deep research.
  • Benchmarks that do not filter parametric memory are partly measuring memorization; adding tool-use tracking and leakage filtering changes the rankings and the conclusions.
  • The pipeline's stages each contribute: the base 30B model scores 40.5%, trajectory SFT raises it to 53.0%, adding text QA SFT brings 56.8%, and GRPO yields 59.3%.
  • The training recipe changes agent behavior, not just accuracy: visual tool invocations rise from roughly 0.10 per task in the 397B baseline to 2.33 in the trained 30B variant, indicating a shift toward balanced multimodal exploration.
  • The RL stage adds a further gain over SFT alone, suggesting that autonomous exploration can be taught beyond the imitation ceiling.

Where Pith is reading between the lines

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

  • If the leakage filter is airtight, the same tool-free rollout check could become a standard sanity test for any benchmark claiming to measure tool-augmented reasoning, not just for video.
  • The stage-wise tool unlocking principle suggests a transferable curriculum for multimodal agents: force competence in one modality's grounding before opening another modality's tools, which could extend to audio, egocentric, or live-streaming research agents.
  • The paper's claim that training methodology outweighs scale is conditional on a single 200-question benchmark; a direct test would be to scale the same pipeline to the 397B base model and evaluate on a larger, independent video-research set.
  • The weaker News-category gain for the 35B variant (8.4 points) compared with the 30B variant (16.6 points) hints that temporally dynamic content may need a dedicated temporal tool, which the paper does not develop.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Video-DeepResearch (Video-DR), a framework for multimodal deep-research agents that operate over continuous video streams rather than static images. The method combines a decoupled perception-exploration pipeline (Select_Keyframe and Crop_Search before text/web search) with a two-stage training recipe: SFT on 7K synthesized video-grounded trajectories plus 7K text-only QA instances, followed by GRPO on a 2K moderate-difficulty RL set. The authors also introduce VIDEODR-BENCH, a human-AI collaboratively annotated benchmark of multi-hop video VQA instances, and claim their Video-DeepResearch-35B-A3B reaches 64.0% average accuracy, surpassing Claude-4.5-Sonnet (59.0%) by 5.0 points. The paper reports substantial gains over the base models, an ablation study, and a tool-usage analysis showing reduced modality bias.

Significance. If the central claims hold, the paper makes a meaningful contribution: it identifies two concrete failure modes of current multimodal agents on video deep research (visual tool aversion and parametric knowledge leakage), and it provides a reproducible training recipe that improves compact open models beyond far larger proprietary systems. The human-AI collaborative benchmark construction and the public code release are also valuable assets to the community. However, the significance is conditional on three load-bearing issues: an unresolved inconsistency in benchmark size, insufficient evidence for the claimed leakage-free and tool-requiring properties of the benchmark, and a circular judge model shared between RL training and evaluation. These issues directly affect whether the headline 5.0-point SOTA improvement is meaningful.

major comments (4)
  1. [Abstract / Section 4.1 / Section 6 / Table 3] The benchmark size is internally inconsistent. The abstract and conclusion state VIDEODR-BENCH comprises 200 instances, while Section 4.1 explicitly says it 'comprises 100 human-annotated VQA pairs.' Table 3's category percentages (29.5 + 22.0 + 18.5 + 14.5 + 12.0 + 3.5 = 100.0) are consistent with 100, not 200. Because the headline 5.0-point SOTA margin is computed on this benchmark, a 100-instance test makes the margin equivalent to roughly 5 questions. Please state the exact number, report per-category counts, and provide confidence intervals or a significance test for the claimed differences.
  2. [Section 3.1 Step 2 / Section 3.4] The claim that 'every question provably requires both visual search and external knowledge reasoning' is not supported by the described filtering. Section 3.1 Step 2 discards a QA pair only if Qwen3.5-397B-A17B answers correctly in any of four tool-free rollouts; Section 3.4 applies a similar filter. This only shows that one particular model could not answer from memory; it says nothing about GPT-5, Claude-4.5-Sonnet, or the trained models, and it does not establish that both visual search and external knowledge are necessary. Additionally, the benchmark is sampled from the same curated video pool used to construct the 30K VQA / 7K training trajectories, and no overlap or contamination analysis is reported. If test items overlap with training data, the reported SOTA gap could reflect memorization rather than tool-augmented exploration. Please provide a quantitative contamination analysis
  3. [Section 3.3 / Section 4.1] The evaluation loop is partially circular. Correctness on VIDEODR-BENCH is judged by Qwen3-VL-30B-A3B-Instruct (Section 4.1), and the same model is used to assign sparse RL rewards in GRPO (Section 3.3). Moreover, Video-DeepResearch-30B-A3B is fine-tuned from exactly Qwen3-VL-30B-A3B-Instruct. This shared judge can systematically favor the trained model's answer distribution. Please evaluate with an independent judge model or human evaluation, and report agreement between judges; at minimum, report results with a different, non-overlapping judge.
  4. [Section 3.1 / Section 3.4 / Figure 2] The benchmark's central property that every item requires both visual search and external knowledge is not validated by any explicit analysis. The leakage filter only checks tool-free answerability; it does not verify that a question is unanswerable with text-only search or with visual-only reasoning. Given the paper's emphasis on modality bias, it is important to show, at least on a sample, that successful answers require both the visual grounding step and the web retrieval step. A decomposition experiment (e.g., text-only access, visual-only access, and full tool set) would substantiate the claim.
minor comments (5)
  1. [Section 4.1 / Table 3 / Table 4] The model name is inconsistent: Table 3 and Table 4 list 'Qwen3.5-397B-A13B,' while Sections 2, 3.1, and 3.2 refer to 'Qwen3.5-397B-A17B.' Please unify the nomenclature.
  2. [Appendix D] The appendix title and text refer to 'VIDEOHUNT,' which appears to be a leftover name; it should be VIDEODR-BENCH throughout.
  3. [Section 3.3] There is an unresolved reference: 'to address the under-utilization of text tools observed in Table??.' Please replace with the correct table number.
  4. [Table 6 / Figure 8] The coordinate scale for select_crop_search is inconsistent: Table 6 specifies normalized [0,1] coordinates, while Figure 8 and the prompt use a 0–1000 scale. Please standardize and make the prompt consistent with the tool specification.
  5. [Table 5] The ablation rows are labeled '4k-SFT' and '14k-SFT+2K-RL,' but the text describes '7K-SFT' and '2K-RL'; the relationship between the 4K and 14K labels and the 7K trajectory count should be clarified.

Circularity Check

2 steps flagged

The benchmark's 'provably requires tools' guarantee is an operational filter from a single model, and the evaluation judge doubles as the GRPO reward model; the SOTA comparison rests on an in-house benchmark with unverified leakage control.

specific steps
  1. self definitional [Sec 3.1 Step 2 (VQA Generation and Verification); benchmark claim in Abstract/Sec 3.4]
    "Post-generation, we rigorously filter out instances prone to parametric memory leakage. Specifically, we conduct four tool-free rollouts for each question; if the agent answers correctly in any attempt, the instance is permanently discarded, guaranteeing that the remaining tasks strictly require external tool utilization."

    The benchmark's headline property—'every question provably requires both visual search and external knowledge reasoning'—is not an external guarantee. 'Requires external tool utilization' is operationally defined as 'Qwen3.5-397B-A17B failed four tool-free rollouts.' This makes the property true by construction only for that model, not for the evaluated models (e.g., Claude-4.5-Sonnet, GPT-5). Questions answerable from another model's parametric memory still pass the filter, so the benchmark does not 'provably' enforce tool-dependence. The later demonstration that tool-augmented models beat tool-free ones is therefore partly a selection artifact: questions the chosen probing model could answer from memory were removed by design.

  2. other [Sec 3.3 (Training/GRPO) and Sec 4.1 (Experimental Setups)]
    "We apply a sparse binary reward, assigning r=1 for correct answers (judged by Qwen3-VL-30B-A3B-Instruct) and r=0 otherwise. ... Correctness is then evaluated by Qwen3-VL-30B-A3B-Instruct, adopting the official judge prompt from Tongyi DeepResearch."

    The 30B model is initialized from Qwen3-VL-30B-A3B-Instruct, the same model that supplies reward labels during GRPO and that later judges correctness on VIDEO-DR-BENCH. The evaluation metric is therefore not independent of the training objective: GRPO explicitly optimizes the policy to produce reports this judge marks 'correct,' and the reported accuracy is then measured by that same judge. The improvements are partly a measure of reward optimization against the judge's idiosyncrasies. This is not a hard logical circularity—the judge compares against reference answers—but it weakens the claim that the gains reflect genuine video deep-research ability rather than alignment to the judge.

full rationale

The central numerical result is not derived from an equation but from a pipeline and an in-house benchmark. The pipeline itself is not circular: SFT on 7K trajectories and GRPO are legitimate training methods, and the ablations show incremental gains. The circularity is in the evaluation loop. First, the benchmark's defining property—that every question requires both visual search and external knowledge—is asserted from a four-rollout tool-free filter using one model, Qwen3.5-397B-A17B. This makes the property relative to that model's parametric memory, so 'provably' is a misdescription; questions answerable by other models from memory are not excluded. Second, correctness on the benchmark is judged by Qwen3-VL-30B-A3B-Instruct, which is also the model that issued GRPO rewards; for the 30B variant it is also the initialization. The evaluation metric is therefore not independent of the training signal, so the reported improvements partly measure optimization against the judge. I do not count the self-citations (Vision-DeepResearch data, Tongyi judge prompt) as circular: they are ordinary reuse. The 100-vs-200 benchmark-size inconsistency and lack of contamination analysis are validity concerns, not circularity. Given the benchmark property reduces by construction and the judge loop is shared, a score of 6 is warranted; the SOTA comparison still contains independent empirical content, so it is not an 8 or 10.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The paper's central claim depends on four hand-set or model-generated quantities (the keyframe similarity threshold, the max keyframes cap, the RL difficulty filter, and the judge model). It also depends on the assumption that model-generated annotations (keyframes, bounding boxes, trajectories) are accurate enough to serve as training labels, and that the leakage filter plus human verification are sufficient to prevent answers from memory. No code, data, or weights are shipped, so the free parameters are not independently verifiable.

free parameters (4)
  • Similarity threshold for keyframe extraction (0.8) = 0.8
    Appendix C: discard consecutive frames with CLIP similarity over 0.8; the threshold is chosen by hand and influences which frames are visible to the agent, and therefore influences the benchmark and training data.
  • Maximum keyframes per video (20) = 20
    Appendix C: caps the input frame count for all videos; it is a hand-set value that limits the agent's visual access.
  • GRPO negative-gradient downsampling probability (20%) = 0.2
    Section 3.3: the advantage of formatting-violation loops is applied with only 20% probability; a hand-chosen training schedule constant, not a scientific parameter per se, but it affects the trained policy.
  • RL dataset difficulty filter (Pass@4 between 0 and 1) = Pass@4 in (0,1)
    Section 3.3: 2K moderate-difficulty dataset retains instances with Pass@4 score strictly between 0 and 1; this is a post hoc selection rule to pick the RL training distribution, which could bias which questions the model learns to solve.
axioms (5)
  • domain assumption CLIP-based inter-frame similarity is a valid proxy for visual redundancy in videos.
    Appendix C: used to select keyframes; if it discards informative frames, the agent loses access to critical entity appearances.
  • domain assumption Qwen3.5-397B-A17B and Qwen3.5-35B-A3B annotations and entity extractions are sufficiently accurate to serve as ground truth for 30K QA pairs and 7K trajectories.
    Section 3.1 and 3.2: these models generate keyframes, bounding boxes, entity names, and trajectory actions; any systematic error in these labels propagates into training and evaluation.
  • domain assumption The four tool-free rollout filter is sufficient to eliminate parametric knowledge leakage.
    Section 3.1 Step 2: a question is discarded if any of four rollouts answers it correctly; this assumes four samples cover the model's memorized knowledge sufficiently. The paper itself identifies knowledge leakage as a major risk, so this is a load-bearing assumption.
  • domain assumption Qwen3-VL-30B-A3B-Instruct is an accurate judge of answer correctness for both RL rewards and final evaluation.
    Sections 3.3 and 4.1: the same model family judges correctness; this assumes judge agreement with human judgment and assumes no bias when the judged model is fine-tuned from the judge model.
  • domain assumption The 100 or 200 benchmark instances are representative of the six video categories and the general video-deep-research task.
    Section 3.4 and Table 2: the benchmark is small (100 or 200, as stated inconsistently) and human-curated; no power analysis or confidence intervals are provided.
invented entities (1)
  • VIDEO-DR-BENCH independent evidence
    purpose: A new 200-instance (or 100-instance, per Section 4.1) benchmark to evaluate video deep research agents.
    The benchmark is a new artifact and could be released publicly, providing an external, falsifiable evaluation instrument. So far, only a description is provided, but the paper claims human-AI annotation and leakage filtering.

pith-pipeline@v1.3.0-daily-deepseek · 18792 in / 9831 out tokens · 78144 ms · 2026-08-05T04:38:36.360211+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent." pith.science (2026). https://pith.science/paper/KC5WV6PW

@misc{pith2026260803979,
  author       = {Pith},
  title        = {Pith review of: Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KC5WV6PW}},
  note         = {Machine review of arXiv:2608.03979}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. Preliminary evaluations reveal two critical bottlenecks in current models: (1) modality bias, where agents bypass visual tools in favor of textual search, and (2) parametric knowledge leakage, where models rely on internal memory rather than genuine tool-augmented execution. To address these challenges, we propose Video-DR, featuring a decoupled perception-exploration pipeline with stage-wise tool unlocking that compels exhaustive cross-frame visual grounding prior to web retrieval. Our framework adopts a two-stage training recipe: supervised fine-tuning followed by Group Relative Policy Optimization (GRPO), enabling autonomous exploration that breaks the imitation-learning ceiling. Furthermore, we curate Video-DR-Bench, a human-AI collaborative benchmark comprising 200 complex, multi-hop VQA instances. Empirical results demonstrate that our Video-DeepResearch-35B-A3B establishes a new state-of-the-art of 64.0% average accuracy, surpassing proprietary Claude-4.5-Sonnet (59.0%) by 5.0 points and significantly outperforming GPT-5 (52.5%) and Gemini 2.5 Pro (57.5%). The 30B-A3B variant achieves 59.3%, competitive with Claude-4.5-Sonnet and demonstrating the effectiveness of our training paradigm even at compact scale. Code: https://github.com/Osilly/Vision-DeepResearch.

Figures

Figures reproduced from arXiv: 2608.03979 by Feng Zhao, Lin Chen, Lionel Z. Wang, Qi Lu, Qingnan Ren, Qingyu Yin, Qisheng Su, Shaohui Lin, Shaosheng Cao, Shiting Huang, Shuang Chen, Tianfei Ren, Wanli Ouyang, Wenxuan Huang, Yao Hu, Yiming Zhao, Yu Zeng, Zehui Chen, Zhen Fang, Zhenfei Yin.

Figure 1
Figure 1. Figure 1: Overview of the VIDEO-DEEPRESEARCH pipeline. Standard visual question-answering metrics are insufficient for measuring an agent’s multi-step strategic execution, creating an urgent need for multi-dimensional evaluation protocols that can ac￾curately quantify long-term reasoning consistency and error-recovery behavior under continuous tem￾poral dynamics. To tackle these challenges, we first conduct an empir… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of VIDEODR-BENCH. The benchmark spans six video domains: Knowledge (29.5%), Entertainment (22.0%), Daily Life (18.5%), Game & Sports (14.5%), News (12.0%), and Others (3.5%). Every instance requires joint visual grounding and multi-hop external knowledge reasoning. ment. To operationalize the spatiotemporal ground￾ing pipeline, the action space A encompasses our defined tools alongside standard we… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of VIDEO-DEEPRESEARCH. Phase I: Raw videos from diverse sources are filtered via rule-based and agent-based stages. Phase II: Keyframes are selected, entities are cropped for visual search, and VQA pairs are synthesized through single- and multi-entity patterns with parametric-leakage filtering. Phase III: Trajectories are constructed via a decoupled perception-exploration pipeline: the agent firs… view at source ↗
Figure 4
Figure 4. Figure 4: The structured evaluation prompt for determining whether a video possesses sufficient semantic depth and [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The structured prompt for multi-frame entity extraction used in Sec. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The structured prompt for knowledge-grounded question generation used in Sec. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The structured evaluation prompt for determining if a deep-research report contains the correct answer. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The prompt for the Deep Research Agent, which has access to all three tools and follows a vision-first [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

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

66 extracted references · 8 canonical work pages

  1. [1]

    arXiv preprint arXiv:2505.22019 , year=

    VRAG-RL: Empower Vision-Perception-Based RAG for Visually Rich Information Understanding via Iterative Reasoning with Reinforcement Learning , author=. arXiv preprint arXiv:2505.22019 , year=

    Pith/arXiv arXiv

  2. [2]

    5-omni technical report , author=

    Qwen3. 5-omni technical report , author=. arXiv preprint arXiv:2604.15804 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2504.07956 , year=

    Vcr-bench: A comprehensive evaluation framework for video chain-of-thought reasoning , author=. arXiv preprint arXiv:2504.07956 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    Pith/arXiv arXiv

  5. [5]

    Introducing Claude Sonnet 4.5 , howpublished =

    work page

  6. [6]

    arXiv preprint arXiv:2601.22060 , year=

    Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models , author=. arXiv preprint arXiv:2601.22060 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2601.06943 , year=

    Watching, Reasoning, and Searching: A Video Deep Research Benchmark on Open Web for Agentic Video Reasoning , author=. arXiv preprint arXiv:2601.06943 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2601.03267 , year=

    Openai gpt-5 system card , author=. arXiv preprint arXiv:2601.03267 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2503.17736 , year=

    V2P-Bench: Evaluating Video-Language Understanding with Visual Prompts for Better Human-Model Interaction , author=. arXiv preprint arXiv:2503.17736 , year=

    arXiv

  10. [10]

    arXiv preprint arXiv:2510.01304 , year=

    Agentic Jigsaw Interaction Learning for Enhancing Visual Perception and Reasoning in Vision-Language Models , author=. arXiv preprint arXiv:2510.01304 , year=

    arXiv

  11. [11]

    arXiv preprint arXiv:2503.06749 , year=

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

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2410.10594 , year=

    Visrag: Vision-based retrieval-augmented generation on multi-modality documents , author=. arXiv preprint arXiv:2410.10594 , year=

    Pith/arXiv arXiv

  13. [13]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Open-domain visual entity recognition: Towards recognizing millions of wikipedia entities , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  14. [14]

    European Conference on Computer Vision , pages=

    Sharegpt4v: Improving large multi-modal models with better captions , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Sharegpt4video: Improving video understanding and generation with better captions , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Advances in Neural Information Processing Systems , volume=

    Are we on the right way for evaluating large vision-language models? , author=. Advances in Neural Information Processing Systems , volume=

  17. [17]

    Mmsearch: Unveiling the potential of large models as multi-modal search engines , author=

  18. [18]

    arXiv preprint arXiv:2601.03193 , year=

    UniCorn: Towards Self-Improving Unified Multimodal Models through Self-Generated Supervision , author=. arXiv preprint arXiv:2601.03193 , year=

    arXiv

  19. [19]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Enhancing Large Vision-Language Models with Ultra-Detailed Image Caption Generation , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  20. [20]

    arXiv preprint arXiv:2511.22134 , year=

    DualVLA: Building a Generalizable Embodied Agent via Partial Decoupling of Reasoning and Action , author=. arXiv preprint arXiv:2511.22134 , year=

    arXiv

  21. [21]

    arXiv preprint arXiv:2506.13977 , year=

    CRITICTOOL: Evaluating Self-Critique Capabilities of Large Language Models in Tool-Calling Error Scenarios , author=. arXiv preprint arXiv:2506.13977 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2509.06945 , year=

    Interleaving reasoning for better text-to-image generation , author=. arXiv preprint arXiv:2509.06945 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2504.05288 , year=

    LiveVQA: Live Visual Knowledge Seeking , author=. arXiv preprint arXiv:2504.05288 , year=

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2506.20670 , year=

    MMSearch-R1: Incentivizing LMMs to Search , author=. arXiv preprint arXiv:2506.20670 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2510.12801 , year=

    Deepmmsearch-r1: Empowering multimodal llms in multimodal web search , author=. arXiv preprint arXiv:2510.12801 , year=

    arXiv

  26. [26]

    arXiv preprint arXiv:2508.05748 , year=

    Webwatcher: Breaking new frontier of vision-language deep research agent , author=. arXiv preprint arXiv:2508.05748 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2511.05271 , year=

    DeepEyesV2: Toward Agentic Multimodal Model , author=. arXiv preprint arXiv:2511.05271 , year=

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2510.24701 , year=

    Tongyi deepresearch technical report , author=. arXiv preprint arXiv:2510.24701 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2505.22648 , year=

    Webdancer: Towards autonomous information seeking agency , author=. arXiv preprint arXiv:2505.22648 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2507.02592 , year=

    WebSailor: Navigating Super-human Reasoning for Web Agent , author=. arXiv preprint arXiv:2507.02592 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2507.15061 , year=

    Webshaper: Agentically data synthesizing via information-seeking formalization , author=. arXiv preprint arXiv:2507.15061 , year=

    Pith/arXiv arXiv

  32. [32]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Simplevqa: Multimodal factuality evaluation for multimodal large language models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  33. [33]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Fvqa: Fact-based visual question answering , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2017 , publisher=

    2017

  34. [34]

    arXiv preprint arXiv:2302.11713 , year=

    Can pre-trained vision and language models answer visual information-seeking questions? , author=. arXiv preprint arXiv:2302.11713 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  36. [36]

    arXiv e-prints , keywords =

    Qwen3-VL Technical Report. arXiv e-prints , keywords =. doi:10.48550/arXiv.2511.21631 , archivePrefix =. 2511.21631 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.21631

  37. [37]

    Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

    Video-llava: Learning united visual representation by alignment before projection , author=. Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

    work page 2024

  38. [38]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Vila: On pre-training for visual language models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  39. [39]

    arXiv preprint arXiv:2410.02713 , year=

    Llava-video: Video instruction tuning with synthetic data , author=. arXiv preprint arXiv:2410.02713 , year=

    Pith/arXiv arXiv

  40. [40]

    thinking with long videos

    Longvt: Incentivizing" thinking with long videos" via native tool calling , author=. arXiv preprint arXiv:2511.20785 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2508.04416 , year=

    Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning , author=. arXiv preprint arXiv:2508.04416 , year=

    Pith/arXiv arXiv

  42. [42]

    arXiv preprint arXiv:2506.13654 , year=

    Ego-r1: Chain-of-tool-thought for ultra-long egocentric video reasoning , author=. arXiv preprint arXiv:2506.13654 , year=

    Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:2512.14870 , year=

    HERBench: A Benchmark for Multi-Evidence Integration in Video Question Answering , author=. arXiv preprint arXiv:2512.14870 , year=

    work page arXiv

  44. [44]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Infinibench: A benchmark for large multi-modal models in long-form movies and tv shows , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2025

  45. [45]

    Advances in Neural Information Processing Systems , volume=

    Longvideobench: A benchmark for long-context interleaved video-language understanding , author=. Advances in Neural Information Processing Systems , volume=

  46. [46]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Lvbench: An extreme long video understanding benchmark , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  47. [47]

    arXiv preprint arXiv:2603.19217 , year=

    Lvomnibench: Pioneering long audio-video understanding evaluation for omnimodal llms , author=. arXiv preprint arXiv:2603.19217 , year=

    arXiv

  48. [48]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Mvbench: A comprehensive multi-modal video understanding benchmark , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  49. [49]

    arXiv preprint arXiv:2510.10689 , year=

    Omnivideobench: Towards audio-visual understanding evaluation for omni mllms , author=. arXiv preprint arXiv:2510.10689 , year=

    arXiv

  50. [50]

    arXiv preprint arXiv:2603.14145 , year=

    MMOU: A Massive Multi-Task Omni Understanding and Reasoning Benchmark for Long and Complex Real-World Videos , author=. arXiv preprint arXiv:2603.14145 , year=

    Pith/arXiv arXiv

  51. [51]

    Advances in Neural Information Processing Systems , volume=

    Sekai: A video dataset towards world exploration , author=. Advances in Neural Information Processing Systems , volume=

    work page

  52. [52]

    arXiv preprint arXiv:2505.21374 , year=

    Video-Holmes: Can MLLM Think Like Holmes for Complex Video Reasoning? , author=. arXiv preprint arXiv:2505.21374 , year=

    Pith/arXiv arXiv

  53. [53]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  54. [54]

    arXiv preprint arXiv:2501.13826 , year=

    Video-mmmu: Evaluating knowledge acquisition from multi-discipline professional videos , author=. arXiv preprint arXiv:2501.13826 , year=

    Pith/arXiv arXiv

  55. [55]

    and Han, Rilyn and Fei-Fei, Li and Xie, Saining , title =

    Yang, Jihan and Yang, Shusheng and Gupta, Anjali W. and Han, Rilyn and Fei-Fei, Li and Xie, Saining , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    work page 2025

  56. [56]

    2026 , eprint=

    WorldSense: Evaluating Real-world Omnimodal Understanding for Multimodal LLMs , author=. 2026 , eprint=

    work page 2026

  57. [57]

    Advances in Neural Information Processing Systems , volume=

    Videoufo: A million-scale user-focused dataset for text-to-video generation , author=. Advances in Neural Information Processing Systems , volume=

    work page

  58. [58]

    arXiv preprint arXiv:2112.09332 , year=

    Webgpt: Browser-assisted question-answering with human feedback , author=. arXiv preprint arXiv:2112.09332 , year=

    Pith/arXiv arXiv

  59. [59]

    arXiv preprint arXiv:2306.02224 , year=

    Auto-gpt for online decision making: Benchmarks and additional opinions , author=. arXiv preprint arXiv:2306.02224 , year=

    Pith/arXiv arXiv

  60. [60]

    arXiv preprint arXiv:2605.05185 , year=

    OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents , author=. arXiv preprint arXiv:2605.05185 , year=

    Pith/arXiv arXiv

  61. [61]

    arXiv preprint arXiv:2602.14234 , year=

    Redsearcher: A scalable and cost-efficient framework for long-horizon search agents , author=. arXiv preprint arXiv:2602.14234 , year=

    arXiv

  62. [62]

    arXiv preprint arXiv:2509.25027 , year=

    Stage: Stable and generalizable grpo for autoregressive image generation , author=. arXiv preprint arXiv:2509.25027 , year=

    arXiv

  63. [63]

    arXiv preprint arXiv:2603.28767 , year=

    Gen-searcher: Reinforcing agentic search for image generation , author=. arXiv preprint arXiv:2603.28767 , year=

    Pith/arXiv arXiv

  64. [64]

    5: Visual Agentic Intelligence , author=

    Kimi K2. 5: Visual Agentic Intelligence , author=. arXiv preprint arXiv:2602.02276 , year=

    Pith/arXiv arXiv

  65. [65]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  66. [66]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021