Pith. sign in

REVIEW 3 major objections 4 minor 39 references

Sparse evidence can suffice: a few OCR/ASR/frame reads reliably decide video misinformation.

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

2026-08-01 16:08 UTC pith:2XEPLOZM

load-bearing objection Genuinely new decoupled evidence-seeking setup, but the verifier-in-the-loop circularity keeps the headline margins from being settled. the 3 major comments →

arxiv 2607.18080 v2 pith:2XEPLOZM submitted 2026-07-20 cs.CV cs.AI

Sparse Evidence Can Suffice: Agentic Evidence Seeking for Multimodal Video Misinformation Detection

Haochen Zhao , Yongxiu Xu , Xinkui Lin , Dong Xie , Jiarui Lu , Yuqi Qian , Yubin Wang , Hongbo Xu
show 1 more author
Gaopeng Gou
This is my paper
classification cs.CV cs.AI
keywords video misinformation detectionevidence-seeking agentsparse evidence sufficiencymultimodal verificationreinforcement learningOCR/ASR evidenceout-of-domain generalizationinterpretability
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 deciding whether a short video is misinformation does not require reading the whole video: a compact package of a few claim-relevant clues from OCR text, speech transcripts, and selected frames is enough. To show this, the authors build SIEVE, in which an evidence-seeking agent learns, via supervised trajectories and reinforcement learning with a frozen verifier as reward, when to read which slot–channel pair and when to stop. On FakeSV and FakeTT, SIEVE outperforms the strongest prior method by about 2.4–2.7 points in accuracy and Macro-F1, and it transfers out of domain to FakeVV without retraining. Its evidence-sufficiency analysis shows that roughly 97% of all eventually-corrected cases are already stably corrected within eight reads, and Macro-F1 saturates around a 12-read budget. A sympathetic reader would care because it reframes the task from holistic video understanding to budgeted evidence acquisition, which is cheaper, faster, and explicitly inspectable.

Core claim

Sparse, claim-relevant evidence can suffice for reliable video misinformation verification. The mechanism: an evidence-seeking agent builds a compact package from slot-structured OCR, ASR, and visual channels, and a frozen verifier judges from that package alone. Trained on teacher trajectories and refined with an evidence-aware RL objective, the agent learns when to stop. On FakeSV and FakeTT, SIEVE scores 92.62/92.41 and 91.64/90.72 ACC/Macro-F1, beating all evaluated baselines. Eight reads already account for about 97% of eventually-corrected cases; going from 12 to 24 reads adds under two points.

What carries the argument

The central objects are the slot-structured evidence space and the decoupled acquisition–verification loop. Each video is divided into temporal slots, each containing up to four channels – OCR text, ASR transcript, context frames (CTX), and key frames (KEY). The agent receives a one-time navigation overview, then issues READ(slot, channel) actions under a read budget until STOP, and only the legally acquired evidence is packed for the Verifier. The mechanism that carries the argument is the budget-constrained interactive loop with a reinforcement-learning reward combining verification outcome, evidence coverage, and interaction penalties; this is what makes the learned policy compact and the

Load-bearing premise

The claim rests on the frozen Verifier being a faithful, non-circular judge of veracity: it was trained only on packages that a third-party model filtered as valid and consistent with the ground-truth label, and it serves both as the RL reward signal and the final test-time evaluator.

What would settle it

Run the same evidence-sufficiency analysis (Figure 3) with the Verifier replaced by an independently trained or human-annotated verifier; if the eight-read/97% saturation disappears or the advantage over exhaustive evidence vanishes, the sparse-sufficiency claim is an artifact of the Verifier's training distribution. Alternatively, build a benchmark where decisive evidence is deliberately spread across many slots and show performance keeps climbing with budget.

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

If this is right

  • Verification no longer needs the whole video: a learned policy that reads on average fewer than eight slot–channel pairs reaches near-saturated Macro-F1 on both in-domain benchmarks.
  • An inference-time budget around 12 reads is enough: extending it to 24 yields only 1.86 (FakeSV) and 0.62 (FakeTT) additional Macro-F1 points.
  • The decoupled design produces an inspectable evidence trail: each READ is logged as (slot, channel), so a claimed verdict can be traced to the exact clues that grounded it.
  • The acquisition policy transfers across domains: the FakeTT-trained checkpoint reaches 73.2 ACC / 74.9 F1 on FakeVV without any FakeVV training, outperforming all baselines under the same no-training protocol.
  • The agent's channel mix is dataset-adaptive (OCR-heavy on FakeSV, balanced OCR/ASR on FakeTT), suggesting the policy learns dataset-specific evidence structure rather than a fixed rule.

Where Pith is reading between the lines

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

  • Extension: the same budgeted acquisition loop may generalize to other video-understanding tasks with sparse decisive evidence (e.g., visual fact-checking, deepfake attribution), but that is our inference, not the paper's claim.
  • Risk: because the frozen Verifier is both reward and evaluator, the reported margins could partly reflect the Verifier's fit to teacher-filtered packages; an independent audit of Verifier labels would test this.
  • Testable extension: one could deploy SIEVE as a pre-filter that reads until confidence is high and passes hard cases to a human fact-checker, estimating how many videos can be resolved within the 8-read budget.
  • Domain boundary: sparse sufficiency was shown on short-video platforms with metadata; on long-form or multi-party videos where decisive evidence is distributed, the hypothesis would need re-testing.

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

3 major / 4 minor

Summary. The paper proposes SIEVE, a framework that decouples evidence acquisition from veracity judgment for multimodal video misinformation detection. An evidence-seeking agent interactively reads OCR, ASR, and visual (CTX/KEY) channels from a slot-based evidence space, then stops to form a compact evidence package; a frozen VLM verifier predicts Fake/Real from the package. The agent is trained with teacher SFT and GRPO using a reward that combines verifier outcome, evidence coverage, and interaction cost. Experiments on FakeSV, FakeTT, and out-of-domain FakeVV report consistent gains over baselines and the paper argues that sparse, claim-relevant evidence packages suffice for reliable verification.

Significance. If the central claim holds, the paper reframes video misinformation detection from holistic multimodal fusion to budgeted, interpretable evidence acquisition, and the explicit evidence trail is a genuine contribution. The controlled ablations (description-only, same-video random evidence, exhaustive evidence) and the evidence-budget analysis are well-designed checks for the sufficiency hypothesis. The main risk is that the frozen verifier serves simultaneously as the RL reward and the final test-time evaluator, while being trained on packages generated with ground-truth-label-conditioned stopping behavior. If the verifier learns label-correlated package structure rather than claim-evidence inconsistency, the reported margins and sufficiency curves are inflated. This concern is distinct from the central hypothesis and could be addressed with additional experiments.

major comments (3)
  1. [Method — Verifier SFT, Eq. (5), Fig. 5, Eq. (6)] The verifier is both the RL reward (Eq. 6) and the final evaluator (Eq. 12), but it is trained only on packages that Kimi K2.5 judges as “valid and consistent with the ground-truth label.” The teacher trajectories that seed verifier training use the ground-truth label as privileged input, and Fig. 5 explicitly instructs label-dependent stopping: for Fake, stop once a falsifier is found; for Real, keep reading until remaining cells look redundant. Package length, channel mix, and item order are therefore correlated with the label. The same-video random ablation controls package size but not this label-correlated distribution. Please provide a content-only assessment (e.g., human or independent oracle judging the same packages without access to acquisition statistics) and/or train/evaluate the verifier on packages from a label-agnostic acquisition policy. If the verifier exploits structura
  2. [Table 1 and Training Details] The paper states that SIEVE results are averaged over three independent runs, but no standard deviations or per-run results are reported. The claimed margins over FakeSV-VLM are +2.40/+2.44 ACC/M-F1 on FakeSV and +2.34/+2.74 on FakeTT; without variance estimates, the reader cannot assess whether these differences are significant. All in-domain baseline numbers are borrowed from FakeSV-VLM without per-run variance. Please report per-run results, standard deviations, and ideally significance tests for SIEVE; for borrowed baselines, clarify whether the same splits and metric implementations were used.
  3. [Can Sparse Evidence Suffice? — Fig. 3(a)] The “stable correction” criterion is retrospective: a sample is counted at step k only if its prediction remains correct for all subsequent evidence prefixes. This is not a decision rule available at inference time, since the future trajectory is unknown. Moreover, the claim that “within eight reads … approximately 97% of all eventually corrected cases” are stably corrected is relative to the eventual-corrected set, not to all Description-only errors, and the absolute sizes of these sets are not reported. Please provide the absolute counts, the distribution of correction steps, and a test-time stopping rule that would realize the claimed sufficiency without oracle knowledge.
minor comments (4)
  1. [Table 2 — Same-video random ablation] Specify how the same-video random packages are sampled, how many random draws are used, and whether the reported numbers are averaged over draws. A single random draw conflates sampling noise with evidence-quality differences.
  2. [Fig. 3(a)] State the size of the initial Description-only error subset and the number of eventually corrected cases on each dataset. The current y-axis “Stable correction rate (%)” is uninformative without these denominators.
  3. [Eqs. (18)-(19) and surrounding text] The KL estimator is described “for completeness” but the formal runs disable KL. If it is not used in the reported experiments, move it to the supplement or remove it to avoid confusion. Also clarify why KL is disabled and whether this choice affects training stability.
  4. [Table 3 and Baselines] For FakeVV, results for Qwen3-VL-8B and GPT-5.5 are evaluated by the authors while the first five baselines are taken from Fact-R1. State the exact prompt, decoding settings, and any filtering used for the two self-evaluated baselines so the comparison is reproducible.

Circularity Check

0 steps flagged

No circular derivation: the final verifier is trained on held-out labels and evaluated against external ground truth; self-cited baselines are not load-bearing.

full rationale

The derivation chain is standard supervised/reward-model training rather than a definitional loop. In Eq. 5, the Verifier is trained on terminal packages D_ver with ground-truth labels; in Eq. 6, the frozen Verifier supplies the correctness signal for RL; in Eq. 12, the same frozen Verifier is used at inference. This is a reward-model pipeline: the reported accuracy is computed against the datasets' held-out ground-truth labels, not against the Verifier's own training criterion, so the prediction is not equal by construction to a fitted input. The teacher's privileged label is used only during rollout and is excluded from stored observations and all Evidence Agent inputs (Method, Teacher trajectory generation; Fig. 5), so the agent is not given y at inference. The label-conditioned stopping hints in Fig. 5 could bias the acquisition distribution, but this is a training-data bias, not a circular reduction. The only self-citations (CA-FVD and FakeSV-VLM as baselines) are used as comparison points, not as justification for the sparse-evidence claim; absolute OOD results on FakeVV and the Exhaustive-evidence ablation provide independent controls. No equation or fitted parameter is renamed as a prediction. Therefore no significant circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

SIEVE's claims rest on hand-chosen budgets and reward coefficients, the reliability of the off-the-shelf extraction stack, distillation from a label-conditioned teacher, a third-party package-quality filter, and quoted baseline numbers. No new physical or conceptual entity is introduced; the 'evidence package' is an organizational structure, not a new entity with independent falsifiable predictions.

free parameters (7)
  • read budget B_read = 24
    Legal READ count cap; hand-chosen. Larger budgets (12->24) yield only +1.86 and +0.62 M-F1 points, so the specific value affects the 'compactness' claim.
  • visual image cap B_img = 24 images
    Cap on accumulated CTX/KEY images in legal read set; hand-chosen.
  • reward outcome weight and base rewards = 0.80 multiplier; +0.20/-1.00
    Eq. (6)/(14): hand-chosen to make correctness primary. The exact balance determines acquisition behavior.
  • evidence coverage Q weights = 0.45/0.25/0.20/0.10
    Eq. (15): hand-tuned to prefer novel text, visual reads, diversity, and volume.
  • interaction penalty caps = P_eff<=0.15, P_beh<=0.30, P_stop=0.03
    Eq. (16)-(17): hand-chosen; removing penalties raises average reads to ~11 and hurts accuracy.
  • GRPO clipping constants = eps_low=0.20, eps_high=0.30, C_dual=3.0, beta=0
    Eq. (19) and KL disabled; standard but arbitrary choices.
  • slot merging and OCR ranking parameters = 0.8s/6s merge thresholds; imp weights 0.25/0.30/0.20/0.15/0.10
    Evidence-space construction hyperparameters; influence what the agent can read. Not fitted to target labels.
axioms (6)
  • domain assumption The Description/title text is a sufficient statement of the claim to be verified; Event, comments, and other metadata are excluded.
    Section 'Problem Formulation' sets c = c_description; if descriptions are misleading, the task changes.
  • domain assumption OCR, ASR, and DINOv2-based visual frames provide reliable, decision-relevant evidence.
    Evidence space construction uses PaddleOCR, Whisper, TransNetV2; noisy extraction would cap agent performance.
  • domain assumption The teacher's label-conditioned stop/read decisions can be distilled into a label-free agent.
    Teacher prompts inject ground-truth hints; the authors filter leakage from observations but the behavioral policy is label-informed.
  • ad hoc to paper Kimi K2.5's assessment that a package is valid and consistent with the label is correct.
    Used to filter verifier training data; no quality analysis of this filter is given.
  • domain assumption Quoted baseline numbers from FakeSV-VLM and Fact-R1 are directly comparable to SIEVE's runs.
    All in-domain baselines in Table 1 are copied from prior work; protocol equivalence is asserted, not demonstrated.
  • domain assumption The frozen verifier's binary judgment is a faithful correctness signal for both RL reward and final evaluation.
    Agent policy is optimized against this verifier; final metrics are measured with the same verifier, so verifier biases propagate into results.

pith-pipeline@v1.3.0-alltime-deepseek · 18774 in / 12727 out tokens · 492672 ms · 2026-08-01T16:08:46.130462+00:00 · methodology

0 comments
read the original abstract

Multimodal video misinformation detection is commonly formulated as a holistic video-understanding task, where the entire video and its associated content are processed and judged in a single pass. However, real-world misinformation often exhibits a sparse and compositional evidence structure: a reliable decision may depend on only a few coupled clues, while most video content contributes limited additional information. Exhaustive multimodal reasoning may therefore introduce substantial redundancy and obscure decisive evidence. This motivates decoupling evidence acquisition from verification: first identifying sparse, decision-relevant clues and then judging veracity based on the acquired evidence. Accordingly, we propose SIEVE, a framework for Sparse Interactive Evidence Verification via Extraction in multimodal video misinformation detection. An evidence-seeking agent actively explores the available multimodal evidence and constructs a compact evidence package, which is then used by a verifier to determine veracity. The agent is trained with supervised evidence-seeking trajectories and an evidence-aware reinforcement learning objective that promotes informative evidence acquisition while discouraging unnecessary or invalid interactions. Experiments on multiple video misinformation benchmarks show that SIEVE consistently outperforms the evaluated baselines and supports reliable verification using compact evidence packages. Moreover, the resulting acquisition process provides an explicit and inspectable evidence trail, improving the transparency and groundedness of multimodal misinformation detection.

Figures

Figures reproduced from arXiv: 2607.18080 by Dong Xie, Gaopeng Gou, Haochen Zhao, Hongbo Xu, Jiarui Lu, Xinkui Lin, Yongxiu Xu, Yubin Wang, Yuqi Qian.

Figure 1
Figure 1. Figure 1: An example illustrating that sparse, claim-relevant [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of the proposed method. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evidence sufficiency on FakeSV and FakeTT. (a) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: System prompt for Teacher trajectory generation. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: User prompt for Teacher trajectory generation. [PITH_FULL_IMAGE:figures/full_fig_p013_6.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

39 extracted references · 1 canonical work pages

  1. [1]

    Proceedings of the 31st ACM International Conference on Multimedia , pages =

    Combating Online Misinformation Videos: Characterization, Detection, and Future Directions , author =. Proceedings of the 31st ACM International Conference on Multimedia , pages =. 2023 , publisher =

    2023

  2. [2]

    Science , volume=

    The spread of true and false news online , author=. Science , volume=. 2018 , doi=

    2018

  3. [3]

    Science , volume=

    The science of fake news , author=. Science , volume=. 2018 , doi=

    2018

  4. [4]

    2023 , doi=

    Qi, Peng and Bu, Yuyan and Cao, Juan and Ji, Wei and Shui, Ruihao and Xiao, Junbin and Wang, Danding and Chua, Tat-Seng , booktitle=. 2023 , doi=

    2023

  5. [5]

    2024 , doi=

    Bu, Yuyan and Sheng, Qiang and Cao, Juan and Qi, Peng and Wang, Danding and Li, Jintao , booktitle=. 2024 , doi=

    2024

  6. [6]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Two Heads Are Better Than One: Improving Fake News Video Detection by Correlating with Neighbors , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=. 2023 , doi=

    2023

  7. [7]

    2025 , doi=

    Wang, JunXi and Wang, Yaxiong and Cheng, Lechao and Zhong, Zhun , booktitle=. 2025 , doi=

    2025

  8. [8]

    Following Clues, Approaching the Truth: Explainable Micro-Video Rumor Detection via

    Hong, Rongpei and Lang, Jian and Xu, Jin and Cheng, Zhangtao and Zhong, Ting and Zhou, Fan , booktitle=. Following Clues, Approaching the Truth: Explainable Micro-Video Rumor Detection via. 2025 , doi=

    2025

  9. [9]

    Advances in Neural Information Processing Systems , volume =

    Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning , author =. Advances in Neural Information Processing Systems , volume =

  10. [10]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle=

  11. [11]

    2602.22963 , archivePrefix=

    Li, Zehao and Yu, Hongwei and Jiang, Hao and Sheng, Qiang and Xu, Yilong and Bi, Baolong and Li, Yang and Yuan, Zhenlong and Cai, Yujun and Wang, Zhaoqi , year=. 2602.22963 , archivePrefix=

    arXiv

  12. [12]

    2023 , eprint=

    The Challenges of Studying Misinformation on Video-Sharing Platforms During Crises and Mass-Convergence Events , author=. 2023 , eprint=

    2023

  13. [13]

    A Multimodal Misinformation Detector for

    Shang, Lanyu and Kou, Ziyi and Zhang, Yang and Wang, Dong , booktitle=. A Multimodal Misinformation Detector for. 2021 , doi=

    2021

  14. [14]

    Proceedings of the 30th ACM International Conference on Information & Knowledge Management , pages=

    Using Topic Modeling and Adversarial Neural Networks for Fake News Video Detection , author=. Proceedings of the 30th ACM International Conference on Information & Knowledge Management , pages=. 2021 , doi=

    2021

  15. [15]

    2023 , doi=

    Liu, Fuxiao and Yacoob, Yaser and Shrivastava, Abhinav , booktitle=. 2023 , doi=

    2023

  16. [16]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2025 , doi=

    2025

  17. [17]

    Proceedings of the ACM on Web Conference 2025 , pages=

    From Predictions to Analyses: Rationale-Augmented Fake News Detection with Large Vision-Language Models , author=. Proceedings of the ACM on Web Conference 2025 , pages=. 2025 , doi=

    2025

  18. [18]

    2026 , doi=

    Wang, Yuxin and Yang, Yang and Zhang, Huaiwen , booktitle=. 2026 , doi=

    2026

  19. [19]

    Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Nip Rumors in the Bud: Retrieval-Guided Topic-Level Adaptation for Test-Time Fake News Video Detection , author=. Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=. 2026 , doi=

    2026

  20. [20]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    From Relevance to Utility: Evidence Retrieval with Feedback for Fact Verification , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=. 2023 , doi=

    2023

  21. [21]

    2018 , doi=

    Thorne, James and Vlachos, Andreas and Christodoulopoulos, Christos and Mittal, Arpit , booktitle=. 2018 , doi=

    2018

  22. [22]

    , title =

    Wu, Zuxuan and Xiong, Caiming and Ma, Chih-Yao and Socher, Richard and Davis, Larry S. , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  23. [23]

    2024 , howpublished =

    GPT-4o mini: advancing cost-efficient intelligence , author =. 2024 , howpublished =

    2024

  24. [24]

    2025 , howpublished =

    Introducing GPT-4.1 in the API , author =. 2025 , howpublished =

    2025

  25. [25]

    arXiv preprint arXiv:2502.13923 , year =

    Qwen2.5-VL Technical Report , author =. arXiv preprint arXiv:2502.13923 , year =

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2412.05271 , year =

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling , author =. arXiv preprint arXiv:2412.05271 , year =

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2411.10442 , year =

    Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization , author =. arXiv preprint arXiv:2411.10442 , year =

    Pith/arXiv arXiv

  28. [28]

    International Conference on Learning Representations , year =

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author =. International Conference on Learning Representations , year =

  29. [29]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding , author =. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =. 2019 , doi =

    2019

  30. [30]

    Advanced Intelligent Computing Technology and Applications , series =

    Consistency-Aware Fake Videos Detection on Short Video Platforms , author =. Advanced Intelligent Computing Technology and Applications , series =. 2025 , publisher =. doi:10.1007/978-981-96-9812-7_17 , note =

    work page doi:10.1007/978-981-96-9812-7_17 2025

  31. [31]

    arXiv preprint arXiv:2008.04838 , year =

    TransNet V2: An Effective Deep Network Architecture for Fast Shot Transition Detection , author =. arXiv preprint arXiv:2008.04838 , year =

    Pith/arXiv arXiv 2008

  32. [32]

    arXiv preprint arXiv:2009.09941 , year =

    PP-OCR: A Practical Ultra Lightweight OCR System , author =. arXiv preprint arXiv:2009.09941 , year =

    Pith/arXiv arXiv 2009

  33. [33]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Robust Speech Recognition via Large-Scale Weak Supervision , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , volume =

    2023

  34. [34]

    Transactions on Machine Learning Research , year =

    DINOv2: Learning Robust Visual Features without Supervision , author =. Transactions on Machine Learning Research , year =

  35. [35]

    2025 , doi=

    Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and others , journal=. 2025 , doi=

    2025

  36. [36]

    2024 , month = dec, url =

    2024

  37. [37]

    2025 , eprint =

    Qwen3-VL Technical Report , author =. 2025 , eprint =

    2025

  38. [38]

    2026 , month = apr, howpublished =

    Introducing GPT-5.5 , author =. 2026 , month = apr, howpublished =

    2026

  39. [39]

    arXiv preprint arXiv:2602.02276 , year =

    Kimi K2.5: Visual Agentic Intelligence , author =. arXiv preprint arXiv:2602.02276 , year =. doi:10.48550/arXiv.2602.02276 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.02276