Pith. sign in

REVIEW 3 major objections 5 minor 56 references

HKD4VLM: A Progressive Hybrid Knowledge Distillation Framework for Robust Multimodal Hallucination and Factuality Detection in VLMs

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A progressive two-stage knowledge-distillation framework lets a 14B vision-language model reach 98.2% F1 at detecting multimodal hallucinations and 98.4% F1 at multimodal fact checking, surpassing GPT-4o and the 72B teacher on both…

desk verdict A credible competition write-up whose central attribution to distillation is confounded by direct supervised fine-tuning; needs a direct-SFT baseline before the mechanism claim holds. read the letter →

arxiv 2506.13038 v2 pith:3TM7JFDO submitted 2025-06-16 cs.CV cs.MM

Zijian Zhang , Xuecheng Wu , Danlei Huang , Siyu Yan , Chong Peng , Xuezhi Cao This is my paper
classification cs.CVcs.MM
keywords multimodalhallucinationdetectionfactualitycheckingknowledgedistillationonlinevision-languagemodelsQwen2.5-VLmulti-tasklearningmappingshift-enhancedinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that a compact vision-language model can be made better than much larger models at two safety-critical verification tasks—detecting hallucinations in AI-generated image descriptions and judging whether textual claims are factual given an image—by distilling knowledge down a ladder of model sizes rather than merely fine-tuning a single big model. The proposed framework HKD4VLM first runs an online distillation cascade from a 72B teacher through a 32B model to a 14B student, then couples all three in a refinement stage, and finally checks answers after shuffling the multiple-choice options. The authors report 98.2% F1 on multimodal hallucination detection and 98.4% F1 on multimodal fact checking, which they claim places first on both competition leaderboards and beats GPT-4o by a wide margin. What is at stake is whether a 14B model, small enough to deploy affordably, can serve as a trustworthy verifier of multimodal content.

What carries the argument

The load-bearing mechanism is logit-space knowledge distillation organized as a capacity ladder: raw pre-softmax outputs of a larger model are matched by a smaller model through a temperature-scaled KL divergence, so the student inherits the teacher's 'dark knowledge' about decisions between options. The framework runs this twice (72B to 32B, then 32B to 14B) in an online mutual-learning mode, then adds Ternary-Coupled Refinement Distillation, in which all three models process the same data and the small model's loss is a weighted mixture of KL terms from the large and medium teachers. At inference, the Mapping Shift-Enhanced Inference strategy permutes the letters attached to options and asks the model to re-answer, which is meant to force it to track semantic content rather than positional cues.

What would settle it

Train the Qwen2.5-VL-14B model directly on the same supervised labels with the same data, schedule, and cross-validation but with no 72B or 32B teacher logits; if that model also reaches the 98% F1 range on the two test sets, the paper's central claim that the distillation cascade is the decisive component is not supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that progressive online distillation transfers fine-grained verification ability across Qwen2.5-VL models of decreasing size: the 72B model is adapted to the task, the 32B model learns from its logits while both are trained together, and the 14B model then learns from the improved 32B; a ternary-coupled refinement adds a dynamically weighted KL loss from both larger models to the small one. With this recipe, the 14B student reaches 98.2% F1 on Multimodal Hallucination Detection and 98.4% F1 on Multimodal Fact Checking, surpassing every baseline in the comparison table, including GPT-4o (66.7% and 71.7% F1 respectively) and the 72B teacher itself (53.1% and 56.1%). The ablation attributes the decisive jump to the pyramid distillation stage, which improves the untuned 14B baseline by 46.1 and 44.5 F1 points, with refinement and option-shuffling inference adding the remaining point.

Load-bearing premise

The reported gain from the pyramid stage is measured against an untuned 14B baseline, and that stage includes cold-start supervised fine-tuning on the task labels, so the interpretation that distillation—rather than direct supervised training—is what produces the improvement is the load-bearing premise.

Editorial extensions

If this is right

  • A 14B model trained by this two-stage distillation can beat both its 72B teacher and external large models on the two tasks, so smaller distillation-based systems are a viable route to responsible multimodal verification.
  • The largest share of the gain comes from the first stage, suggesting the cascade design—not the refinement or inference tricks alone—is what transfers task knowledge.
  • The ternary refinement stage adds consistent small gains, and a smaller weighting factor that leans on the medium model as a bridge works better than leaning on the large model.
  • Option-shuffling inference improves robustness against positional bias, a benefit that should transfer to other multiple-choice multimodal tasks.
  • Training one model jointly on hallucination detection and fact checking under a multi-task objective yields synergistic gains, supported by five-fold cross-validation and data augmentation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's ablation does not include a direct-supervision-only condition, so a natural follow-up experiment would be to train the 14B model on the same labels without any teacher logits; the reported +46-point jump would then separate the contribution of distillation from the contribution of direct supervision.
  • The test sets contain 1,000 images each and the tasks use a multiple-choice format; evaluating the distilled model on open-ended generation and on standard hallucination benchmarks would clarify whether the near-ceiling scores reflect a general verification skill or a format-specific gain.
  • If the cascade works mainly by providing a strong teacher-logit signal, a simpler two-model setup (72B teacher directly to 14B student) may capture most of the benefit at lower training cost, since the ternary stage adds less than one F1 point.
  • The option-permutation trick could serve as a general robustness audit for positional bias in any multiple-choice vision-language evaluation, not just the two competition tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper describes HKD4VLM, a two-stage knowledge-distillation framework for multimodal hallucination detection (MHD) and multimodal factuality checking (MFC). Stage 1, Pyramid-like Progressive Online Distillation, co-trains Qwen2.5-VL models of sizes 72B, 32B, and 14B using KL-divergence distillation losses plus cross-entropy against ground-truth labels. Stage 2, Ternary-Coupled Refinement Distillation, adds a dynamically weighted KL loss from the large and medium models to the small model. Inference uses Mapping Shift-Enhanced Inference (MSEI), which permutes answer options and re-evaluates. The authors report 98.2% F1 on MHD and 98.4% F1 on MFC, ranking first on both challenge leaderboards; the ablation in Table 4 attributes gains of +46.1 and +44.5 F1 to the pyramid stage. No code or leaderboard link is provided.

Significance. If the results withstand scrutiny, the paper would demonstrate a practical recipe for obtaining a compact, high-accuracy VLM for hallucination and factuality detection, an important safety-critical application. The claimed first-place leaderboard positions are notable, and the multi-task, multi-teacher distillation design is well motivated. However, the central mechanistic claim that distillation—rather than direct supervised fine-tuning on the challenge labels—drives the dramatic gains is not tested by the current ablation design. The paper also lacks error bars and code, so the magnitude and reproducibility of the reported increments cannot be assessed. The significance is therefore conditional on the ablations being completed.

major comments (3)
  1. [Table 4; Sec. 3.2, Eq. (3)] The ablation baseline for the pyramid stage is the untuned Qwen2.5-VL-14B (51.2 MHD / 53.0 MFC in Table 3). The 'Pyramid-like Progressive Online Distillation' condition, however, includes supervised fine-tuning: Eq. (3) contains L_CE(y, z_SM) and L_CE(y, z_SS), and the 72B teacher is cold-start SFT on the combined official dataset. The reported +46.1/+44.5 F1 gains therefore conflate direct supervised training on the challenge labels (via LoRA, multi-task formatting, augmentation, and five-fold training) with knowledge transfer from teacher logits. To support the abstract's claim that a smaller distilled VLM outperforms a larger directly-tuned VLM, the authors must report a direct-SFT-14B baseline trained with the same data, augmentation, and folds but without the distillation losses, and preferably also direct-SFT-32B and direct-SFT-72B conditions. Without these controls, the central 'distillation helps' claim is unsupported.
  2. [Sec. 3.4; Table 4] The paper states that five-fold cross-validation is used and final scores are averaged over the five test folds, but Table 4 reports only single F1 values with no standard deviations, per-fold results, or significance tests. The incremental gains attributed to TCRD (+0.6/+0.5) and MSEI (+0.3/+0.4) are small enough that they could plausibly fall within fold-to-fold or training-seed variance. Reporting variance is necessary to support the progressive-improvement narrative in Sec. 4.4 and to allow readers to judge whether the later stages add reliable gains.
  3. [Sec. 3.3; Table 4] The Mapping Shift-Enhanced Inference strategy is underspecified. The text describes permuting option contents and re-evaluating, but it does not state how the original and permuted evaluations are combined into a final answer, how many permutations are applied per question, whether the strategy is used only at test time or also during training, or its computational overhead. Consequently, the +0.3/+0.4 F1 gain attributed to MSEI is not reproducible from the description, and the robustness claim cannot be verified.
minor comments (5)
  1. [Sec. 4.4 / Sec. 4.1] Typographical errors: 'devrise' should be 'diverse' in Sec. 4.4, and 'empoloy' should be 'employ' in Sec. 4.1; also 'first palce' appears in the Introduction.
  2. [Eq. (4) and surrounding text] The description of the gamma weighting is reversed: a smaller gamma increases the weight on L_KD(z_SM, z_SS), which is distillation from the medium model to the small model, not 'distillation from the small model to the medium model' as stated. The equation itself is clear, but the text should be corrected.
  3. [Sec. 4.1] Training details are too sparse for reproducibility: the paper gives an initial learning rate and cosine schedule, but not the LoRA rank, alpha, target modules, batch size, number of epochs, sequence length, or the values of alpha, beta, and tau used in Eqs. (3) and (4). The phrase 'other settings keep consistent with the default environment' is not sufficient for a methods paper.
  4. [Table 3] The comparison in Table 3 is between fine-tuned HKD4VLM and out-of-the-box or differently-tuned baselines; the paper should explicitly acknowledge that this is a challenge-leaderboard comparison rather than a controlled comparison, since GPT-4o and the other listed VLMs are not fine-tuned on the same challenge data.
  5. [Abstract / Introduction] The claim of ranking first on both leaderboards is not externally verifiable from the manuscript; a leaderboard URL or citation of the challenge results page should be provided.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the only concern is a confounded ablation, which is a validity issue rather than a circular step.

full rationale

This is an empirical systems paper with no formal derivation whose predictions reduce to their inputs by construction. The central results are benchmark scores on official testbeds, and the proposed HKD4VLM is evaluated externally against strong baselines (GPT-4o, InternVL-3-78B, etc.), so the main claim is not a renamed input or a fitted parameter called a prediction. The self-citations present (e.g., [37], [40], [41], [42], [47]) are background, protocol, or related-work references and are not load-bearing for the framework's effectiveness. The only notable issue is an experimental-design confound: in Table 4, the +46.1/+44.5 F1 gain attributed to 'Pyramid-like Progressive Online Distillation' is measured against the untuned Qwen-2.5-VL-14B baseline, while Eq. (3) shows that this component includes direct supervised cross-entropy losses L_CE(y, z_SM) and L_CE(y, z_SS) on the ground-truth labels. The ablation therefore does not isolate distillation from ordinary supervised fine-tuning, and the claim that distillation is the critical component is not uniquely established. However, this is a confound in the supporting evidence, not a circularity: the paper does not define the target result in terms of the model's own outputs, nor does it cite its own prior work to forbid alternatives. Accordingly, no circular step is identified, and the paper receives a low score reflecting a minor concern rather than actual circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The framework is an empirical system without a formal derivation. The main 'free parameters' are distillation weighting hyperparameters (alpha, beta, gamma, tau) whose values are mostly undisclosed. The core assumptions are that teacher logits carry transferable knowledge, that five-fold validation on the challenge training set is representative, that option-label shuffling preserves semantics, and that the competition leaderboard is reliable. No new physical entities are introduced.

free parameters (4)
  • gamma (TCRD dynamic weighting factor) = 0.10
    Tuned on validation across 0.75, 0.50, 0.25, 0.10; the best value is reported in Table 5 and used in the final method.
  • alpha (stage-1 distillation weight)
    Balances cross-entropy and KL loss for the medium student in Eq. (3); value is not disclosed.
  • beta (stage-2 distillation weight)
    Balances cross-entropy and KL loss for the small student in Eq. (3); value is not disclosed.
  • temperature tau
    Used in the KL distillation loss in Eq. (2) to soften logits; value is not disclosed.
assumptions (4)
  • domain assumption Teacher logits encode transferable 'dark knowledge' that helps the student beyond hard labels.
    Core premise of the distillation stage in Section 3.2; not independently verified in the paper.
  • domain assumption Five-fold cross-validation on the challenge training set is representative of the hidden test set.
    The method reports averaged metrics over five folds (Section 3.4) and assumes this transfers to the leaderboard.
  • domain assumption Shuffling option labels preserves the semantic content of the options.
    The mapping shift-enhanced inference (Section 3.3) relies on this invariance to improve robustness.
  • domain assumption The challenge leaderboard scores are reliable and not affected by data leakage.
    The headline SOTA results are taken from an external competition without independent verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HKD4VLM: A Progressive Hybrid Knowledge Distillation Framework for Robust Multimodal Hallucination and Factuality Detection in VLMs." pith.science (2026). https://pith.science/paper/3TM7JFDO

@misc{pith2026250613038,
  author       = {Pith},
  title        = {Pith review of: HKD4VLM: A Progressive Hybrid Knowledge Distillation Framework for Robust Multimodal Hallucination and Factuality Detection in VLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TM7JFDO}},
  note         = {Machine review of arXiv:2506.13038}
}
read the original abstract

Driven by the rapid progress in vision-language models (VLMs), the responsible behavior of large-scale multimodal models has become a prominent research area, particularly focusing on hallucination detection and factuality checking. In this paper, we present the solution for the two tracks of Responsible AI challenge. Inspirations from the general domain demonstrate that a smaller distilled VLM can often outperform a larger VLM that is directly tuned on downstream tasks, while achieving higher efficiency. We thus jointly tackle two tasks from the perspective of knowledge distillation and propose a progressive hybrid knowledge distillation framework termed HKD4VLM. Specifically, the overall framework can be decomposed into Pyramid-like Progressive Online Distillation and Ternary-Coupled Refinement Distillation, hierarchically moving from coarse-grained knowledge alignment to fine-grained refinement. Besides, we further introduce the mapping shift-enhanced inference and diverse augmentation strategies to enhance model performance and robustness. Extensive experimental results demonstrate the effectiveness of our HKD4VLM. Ablation studies provide insights into the critical design choices driving performance gains.

Figures

Figures reproduced from arXiv: 2506.13038 by the authors.

Figure 1
Figure 1. The overall illustrations of our proposed HKD4VLM [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The illustrations of online knowledge distillation. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overall illustrations of our proposed mapping [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization comparisons of HKD4VLM and advanced VLMs in terms of final answer and logits on two tasks. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The comparisons of loss curves in the ternary [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 19 canonical work pages

  1. [1]

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al

  2. [2]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision- language model with versatile abilities.arXiv preprint arXiv:2308.12966(2023)

    arXiv 2023

  3. [3]

    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)

    arXiv 2025

  4. [4]

    Alimohammad Beigi, Bohan Jiang, Dawei Li, Tharindu Kumarage, Zhen Tan, Pouya Shaeri, and Huan Liu. 2024. Lrq-fact: Llm-generated relevant questions for multimodal fact-checking.arXiv preprint arXiv:2410.04616(2024)

    arXiv 2024

  5. [5]

    Kilian Carolan, Laura Fennelly, and Alan F Smeaton. 2024. A review of multi- modal large language and vision models.arXiv preprint arXiv:2404.01322(2024)

    arXiv 2024

  6. [6]

    Recep Firat Cekinel, Pinar Karagoz, and Cagri Coltekin. 2024. Multimodal Fact- Checking with Vision Language Models: A Probing Classifier based Solution with Embedding Strategies.arXiv preprint arXiv:2412.05155(2024)

    work page Pith review arXiv 2024

  7. [7]

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al . 2024. A survey on evaluation of large language models.ACM transactions on intelligent systems and technology15, 3 (2024), 1–45

    2024

  8. [8]

    Zhe Chen, Jiannan Wu, et al. 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 24185–24198

    work page 2024

Show all 56 references
  1. [9]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, et al . 2023. Instructblip: Towards general-purpose vision- language models with instruction tuning.arXiv preprint:2305.06500(2023)

    2023 arXiv

  2. [10]

    Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdh- ery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. 2023. Palm-e: An embodied multimodal language model.arXiv preprint arXiv:2303.03378(2023)

    2023 arXiv

  3. [11]

    Tobias Falke, Leonardo FR Ribeiro, Prasetya Ajie Utama, Ido Dagan, and Iryna Gurevych. 2019. Ranking generated summaries by correctness: An interesting but challenging application for natural language inference. InProceedings of the 57th annual meeting of the association for c...

    2019

  4. [12]

    Yunfan Gao, Yun Xiong, Yijie Zhong, Yuxi Bi, Ming Xue, and Haofen Wang

  5. [13]

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. 2024. Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680 (2024)

    2024 arXiv

  6. [14]

    Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. 2024. A comprehensive survey of retrieval-augmented generation (rag): Evolution, current landscape and future directions.arXiv preprint arXiv:2410.12837(2024)

    2024 arXiv

  7. [15]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531(2015)

    2015 arXiv

  8. [16]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.ICLR1, 2 (2022), 3

    2022

  9. [17]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation.ACM computing surveys55, 12 (2023), 1–38

    2023

  10. [18]

    Hugo Laurençon, Lucile Saulnier, et al. 2024. Obelics: An open web-scale filtered dataset of interleaved image-text documents.Advances in Neural Information Processing Systems36 (2024)

    2024

  11. [19]

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen

  12. [20]

    Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang

  13. [21]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024. Visual instruc- tion tuning.Advances in neural information processing systems36 (2024)

    2024

  14. [22]

    Lihui Liu, Houxiang Ji, Jiejun Xu, and Hanghang Tong. 2022. Comparative Rea- soning for Knowledge Graph Fact Checking. In2022 IEEE International Conference on Big Data (Big Data)

    2022

  15. [23]

    Ye Liu, Jiajun Zhu, Xukai Liu, Haoyu Tang, Yanghai Zhang, Kai Zhang, Xiaofang Zhou, and Enhong Chen. 2025. Detect, Investigate, Judge and Determine: A Knowledge-guided Framework for Few-shot Fake News Detection

    2025

  16. [24]

    Mitigating hallucination in large multi-modal models via robust instruction tuning.arXiv preprint arXiv:2306.14565(2023)

    2023 arXiv

  17. [25]

    Humza Naveed, Asad Ullah Khan, et al. 2023. A comprehensive overview of large language models.arXiv preprint arXiv:2307.06435(2023)

    2023 arXiv

  18. [26]

    OpenAI. 2024. Hello GPT-4o. https://openai.com/index/hello-gpt-4o/

    2024

  19. [27]

    A Paszke. 2019. Pytorch: An imperative style, high-performance deep learning library.arXiv preprint arXiv:1912.01703(2019)

    2019 arXiv

  20. [28]

    Holy Lovenia, Wenliang Dai, et al. 2023. Negative object presence evaluation (nope) to measure object hallucination in vision-language models.arXiv preprint arXiv:2310.05338(2023)

    2023 arXiv

  21. [29]

    Liyan Tang, Philippe Laban, and Greg Durrett. 2024. MiniCheck: Efficient Fact- Checking of LLMs on Grounding Documents. InProceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    2024

  22. [30]

    GLM Team, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. 2024. Chatglm: A fam- ily of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793(2024)

    2024 arXiv

  23. [31]

    Qwen Team. 2024. Qvq: To see the world with wisdom

    2024

  24. [32]

    Tanik Saikh, Arkadipta De, Asif Ekbal, and Pushpak Bhattacharyya. 2020. A Deep Learning Approach for Automatic Detection of Fake News.arXiv preprint arXiv:2005.04938(2020)

    2020 arXiv

  25. [33]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

    2023 arXiv

  26. [34]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191(2024)

    2024 arXiv

  27. [35]

    Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. 2022. Effi- cientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning.arXiv preprint arXiv:2210.07795(2022)

    2022 arXiv

  28. [36]

    Tencent. 2025. Hunyuan-Vision. https://hunyuan.tencent.com

    2025

  29. [37]

    Yifan Wang, Xuecheng Wu, Jia Zhang, et al. 2024. Building Robust Video-Level Deepfake Detection via Audio-Visual Local-Global Interactions. InProceedings of the 32nd ACM International Conference on Multimedia. 11370–11376

    2024

  30. [38]

    Yifan Wang, Jian Zhao, Zhaoxin Fan, Xin Zhang, Xuecheng Wu, Yudian Zhang, et al. 2025. JTD-UAV: MLLM-Enhanced Joint Tracking and Description Frame- work for Anti-UAV Systems. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR). 1633–1644

    2025

  31. [39]

    Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and Philip S Yu

  32. [40]

    Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. 2023. Cogvlm: Visual expert for pretrained language models.arXiv preprint arXiv:2311.03079(2023)

    2023 arXiv

  33. [41]

    Xuecheng Wu, Heli Sun, Yifan Wang, Jiayu Nie, Jie Zhang, Yabing Wang, Junxiao Xue, and Liang He. 2025. AVF-MAE++: Scaling Affective Video Facial Masked Autoencoders via Efficient Audio-Visual Self-Supervised Learning. InProceedings of the Computer Vision and Pattern Recognitio...

    2025

  34. [42]

    Xuecheng Wu, Heli Sun, Junxiao Xue, Jiayu Nie, Xiangyan Kong, Ruofan Zhai, and Liang He. 2023. Towards Emotion Analysis in Short-form Videos: A Large- Scale Dataset and Baseline.arXiv e-prints(2023), arXiv–2311

    2023

  35. [43]

    Yang Yang, Lei Zheng, Jiawei Zhang, Qingcai Cui, Zhoujun Li, and Philip S. Yu

  36. [44]

    In2023 IEEE International Conference on Big Data (BigData)

    Multimodal large language models: A survey. In2023 IEEE International Conference on Big Data (BigData). IEEE, 2247–2256

  37. [45]

    Xuecheng Wu, Jiaxing Liu, Danlei Huang, Xiaoyu Li, Yifan Wang, Chen Chen, Liya Ma, Xuezhi Cao, and Junxiao Xue. 2025. ViC-Bench: Benchmarking Visual- Interleaved Chain-of-Thought Capability in MLLMs with Free-Style Intermediate State Representations.arXiv preprint arXiv:2505.1...

    2025

  38. [46]

    Zaiwei Zhang, Gregory P Meyer, Zhichao Lu, Ashish Shrivastava, Avinash Ravichandran, and Eric M Wolff. 2024. VLM-KD: Knowledge Distillation from VLM for Long-Tail Visual Recognition.arXiv preprint arXiv:2408.16930(2024)

    2024 arXiv

  39. [47]

    Zijian Zhang, Xuhui Zheng, Xuecheng Wu, Chong Peng, and Xuezhi Cao. 2025. TokenFocus-VQA: Enhancing Text-to-Image Alignment with Position-Aware Focus and Multi-Perspective Aggregations on LVLMs. InProceedings of the Com- puter Vision and Pattern Recognition Conference (CVPR) W...

    2025

  40. [48]

    Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yunlin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, Wenmeng Zhou, and Yingda Chen. 2024. SWIFT:A Scalable lightWeight Infrastructure for Fine-Tuning. arXiv:2408.05517 [cs.CL] https://arxiv.org/abs/2408.05517

    2024 arXiv

  41. [49]

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models.arXiv preprint arXiv:2304.10592(2023)

    2023 arXiv

  42. [50]

    Yu-Chu Yu, Chi-Pin Huang, Jr-Jen Chen, Kai-Po Chang, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. 2024. Select and Distill: Selective Dual-Teacher Knowledge Transfer for Continual Learning on Vision-Language Models

    2024

  43. [51]

    Hospedales, and Huchuan Lu

    Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. 2017. Deep Mutual Learning.arXiv preprint arXiv:1706.00384(2017)

    2017 arXiv

  44. [56]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, et al. 2025. InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models. arXiv preprint arXiv:2504.10479(2025)

    2025 arXiv

  45. [2018]

    TI-CNN: Convolutional Neural Networks for Fake News Detection.arXiv preprint arXiv:1806.00749(2018)

    2018 arXiv

  46. [2022]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems35 (2022), 23716–23736

    2022

  47. [2023]

    Evaluating object hallucination in large vision-language models.arXiv preprint arXiv:2305.10355(2023)

    2023 arXiv

  48. [2025]

    Synergizing rag and reasoning: A systematic review.arXiv preprint arXiv:2504.15909(2025)

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.