Pith. sign in

REVIEW 33 references

The paper establishes that caption quality splits into coverage and precision, and that each drives a different downstream task.

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

2026-08-04 03:31 UTC pith:KEIOOOET

arxiv 2608.02589 v1 pith:KEIOOOET submitted 2026-08-03 cs.CV

CAPEval: A Decoupled Caption Evaluation across Understanding and Generation

Zhipeng Liu , Haochen Wang , Zhaoxiang Zhang This is my paper
classification cs.CV
keywords caption evaluationcoverageprecisionvision-language understandingtext-to-image generationbenchmarktask-dependent dissociationcaption data curation
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 argues that caption quality is not a single scalar but two independent properties: coverage (how much of an image's ground-truth content a caption mentions) and precision (how many of a caption's claims are actually correct). Previous caption metrics and benchmarks collapse these into one score, obscuring why a caption is useful. CAPEval measures both via dense human-verified checklists judged by a large language model, then uses captions from ten different captioners as the only variable in controlled training runs of vision-language models and text-to-image models. The central finding is a task-dependent dissociation: coverage is the statistically significant predictor of understanding performance, while precision is the statistically significant predictor of generation performance. If correct, this means caption data should be curated differently depending on the target task.

Core claim

On its own terms, CAPEval finds that the two axes it defines behave differently across downstream objectives. Regressing aggregate understanding and generation scores on coverage and precision across ten captioners, coverage carries a significant positive coefficient in both understanding pipelines (beta_C = +0.118 and +0.215) while precision does not; for generation, precision is the only significant predictor (beta_P = +0.189 and +0.235) while coverage is not. The same pattern appears inside a single captioner family, where the smallest model's higher coverage coincides with better understanding and a mid-size model's higher precision coincides with better generation. The paper concludes t

What carries the argument

The central mechanism is the (Coverage, Precision) pair defined over atomic checklist items. Each of 300 images has a human-written ground-truth caption decomposed into about fifty binary, human-verified checklist items across eight semantic dimensions (instance, attribute, relation, image, text, human, UI, world knowledge). A large language model judge reads a candidate caption and marks each item as correctly mentioned, wrongly mentioned, or not mentioned; Coverage is the fraction of items the caption attempts, and Precision is the fraction of attempts that are correct. These two numbers replace a single scalar caption score. The other load-bearing component is the controlled training prot

Load-bearing premise

Everything rests on the automatic judge's yes/no/not-mentioned labels being accurate and unbiased across captioners; if the judge systematically favors certain caption styles, lengths, or model families, both Coverage and Precision are biased, and the observed dissociation between understanding and generation could be an artifact of judge bias rather than a property of captions.

What would settle it

Swap the judge model (or have humans relabel a random subset of the 300 images) and recompute Coverage and Precision for the ten captioners; if the regression dissociation between understanding and generation changes sign or loses statistical significance, the central claim is judge-dependent rather than a property of captions.

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

If this is right

  • Single-scalar caption metrics are insufficient: two captions with the same aggregate score can have opposite coverage-precision profiles and thus opposite utility for a given task.
  • Multimodal understanding systems should be trained with captions chosen for broad semantic coverage, even if that means tolerating some imprecision in the captions.
  • Text-to-image generation systems should be trained with captions chosen for high factual precision, even if that means narrower coverage.
  • Captioner scale is not a reliable proxy for downstream utility: within one family, the smallest variant outperformed the largest for understanding, and a mid-size variant outperformed the largest for generation.
  • Hallucination robustness in understanding models is an exception: it correlates more strongly with caption precision than with coverage, so precision-oriented curation may help reduce hallucination.

Where Pith is reading between the lines

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

  • A direct, untested consequence is a practical recipe for caption data curation: generate broadly, then filter for precision to serve generation, or keep breadth for understanding - a two-stage pipeline CAPEval itself does not spell out.
  • Because CAPEval scores depend on a single judge model, a natural extension is to average scores over multiple judges or add human spot-checking; the regression slopes observed here are likely to shift if judge agreement is low.
  • The same decoupled evaluation could be applied to other text-conditioned tasks, such as text-to-video or editing, where the understanding-generation asymmetry may or may not replicate.
  • If the dissociation is causal, reinforcement-learning pipelines that reward caption quality should use a task-specific reward - coverage for VLM pretraining, precision for T2I - rather than a combined scalar.

Editorial analysis

A structured set of objections, weighed in public.

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

Circularity Check

0 steps flagged

No circularity found: CAPEval scores and downstream U/G are measured in independent pipelines, with no fitted quantity renamed as a prediction.

full rationale

CAPEval's derivation chain does not contain a circular step. Coverage and Precision are defined from Qwen2.5-72B judgments over human-verified checklists (Eqs. 1-2), while the downstream understanding and generation scores come from separate, controlled VLM/T2I training and evaluation pipelines. The OLS regressions in Eq. 3 then relate these independently measured quantities across ten captioners. No equation defines C or P in terms of U or G, no fitted parameter is subsequently 'predicted' from the same fitted values, and the central dissociation is an empirical correlation rather than an identity. The lack of human validation of judge labels, the small n=10 regression, and the absence of held-out validation are legitimate measurement and statistical-strength concerns, but they are not circularity. The paper contains self-citations in the background and related work, but they are not load-bearing for the benchmark construction or the dissociation result, which rest on new checklists, new downstream experiments, and externally available models and benchmarks. Therefore the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physical entities are postulated. The paper's contributions are measurement constructs (Coverage, Precision, CAPEval) and an empirical relation; these are measured rather than invented. The central empirical claims rest on the judge's accuracy and on an unspecified benchmark normalization, both of which are unverified assumptions.

free parameters (3)
  • OLS coefficients β_C and β_P per pipeline = SigLIP: +0.118/−0.034; CLIP: +0.215/−0.049; SD3.5M: +0.028/+0.189; Qwen-Image: −0.076/+0.235
    Fitted in-sample on 10 captioners and used as the main evidence for the dissociation; no out-of-sample validation or confidence interval is reported.
  • Benchmark score normalization mapping = unspecified
    Appendix A.5 says raw benchmark scores are 'mapped to a common 0–100 scale' without giving the mapping; different choices can change the aggregate U and G values.
  • Qwen2.5-72B judge classification thresholds
    The judge outputs yes/no/not-mentioned for each checklist item; no prompt details, temperature, or human-agreement calibration are reported, so the effective threshold is a free modeling choice.
axioms (5)
  • domain assumption Atomic checklist items are an unbiased and complete representation of the visual content an image caption should mention.
    Load-bearing for C and P (Sec. 3.2): if the checklists omit common facts captions express, Coverage and Precision are systematically distorted.
  • domain assumption Qwen2.5-72B judge labels are accurate and unbiased across all captioners.
    C and P are fully determined by the judge (Sec. 3.3); no human reliability check of judge labels is reported.
  • domain assumption Captions from each captioner are the only varying factor in downstream training, with no hidden confounds such as caption style, model family, or token-length filtering.
    The controlled protocol is stated in Sec. 3.4, but captioner identity is not randomized, so model-family or stylistic confounds remain.
  • standard math OLS linearity, residual independence, and approximate normality hold for regressions with 10 captioners.
    Used to compute p-values in Sec. 4.2; no residual diagnostics, collinearity checks, or correction for multiple comparisons are reported.
  • ad hoc to paper The unstated normalization of benchmark scores is monotone and comparable across benchmarks.
    Appendix A.5 states scores are normalized to 0–100 but does not define the mapping; arbitrary choices can affect the aggregate U/G and the regression conclusions.

pith-pipeline@v1.3.0-daily-deepseek · 22391 in / 15073 out tokens · 213854 ms · 2026-08-04T03:31:17.826064+00:00 · methodology

0 comments
read the original abstract

Captions serve as a primary supervision signal for both multimodal understanding and text-to-image generation. However, previous evaluations treat the caption quality as a single scalar objective, which conflates two distinct properties: (1) how much visual information a caption covers and (2) how reliably the image supports its stated claims. To this end, we design a decoupled caption evaluation benchmark, CAPEval (Coverage And Precision Evaluation), with human-written ground-truth captions and human-verified atomic checklist items. Specifically, CAPEval decomposes caption quality into Coverage and Precision. The former quantifies how thoroughly a caption covers ground-truth factual content, while the latter reflects the factual correctness rate of all claims expressed in the caption. We select 10 captioners and further conduct controlled downstream end-to-end experiments with them from four model families, where the caption source is the only variable. Empirically, we find a consistent task-dependent dissociation: Coverage serves as the stronger correlate for understanding performance, whereas Precision acts as the dominant predictor for generation performance. This decoupled evaluation paradigm not only delivers a more fine-grained diagnosis of caption quality, but also offers actionable guidance for selecting and optimizing captioners tailored to different downstream tasks.

Figures

Figures reproduced from arXiv: 2608.02589 by Haochen Wang, Zhaoxiang Zhang, Zhipeng Liu.

Figure 1
Figure 1. Figure 1: Overview of CAPEval. (a) CAPEval constructs fine-grained semantic checklists from high-resolution (up to 8K) images, long-form ground-truth captions, and dense atomic checklist items spanning eight semantic dimensions. (b) Each candidate caption is evaluated against the atomic checklist items by a judge model to obtain Coverage (C) and Precision (P). (c) Captions produced by different captioners are then u… view at source ↗
Figure 2
Figure 2. Figure 2: Relating CAPEval scores to downstream performance reveals a task-dependent asymme [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: CAPEval image collection and annotation design. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of CAPEval with existing benchmarks across [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Coverage (C) and Precision (P) scores of multimodal models on CAPEval. The dashed horizontal line separates proprietary models from open-source models. Captioners exhibit diverse and non-monotonically aligned Coverage–Precision profiles, showing that the two dimensions cap￾ture distinct properties of caption quality. 4 EXPERIMENTS 4.1 CAPEVAL METRICS BETTER EXPLAIN DOWNSTREAM UTILITY THAN SCALE CAPEval Rev… view at source ↗
Figure 6
Figure 6. Figure 6: Regression analysis across four downstream pipelines. Row labels show the fitted equa [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Regression analysis on hallucination score shows the fitted equation for the average score of CLIP-Vicuna and SigLIP-Qwen3. For hallucination, Precision is more significant than Coverage. 5 CONCLUSION In this paper, we introduced CAPEval, a framework that decomposes caption quality into two distinct axes: (1) Coverage, measuring how much of the visual content a caption attempts to describe, and (2) Precisi… view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of image resolutions (height [PITH_FULL_IMAGE:figures/full_fig_p018_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

33 extracted references · 26 linked inside Pith

  1. [4]

    19 Table 4: Normalized understanding and generation benchmark scores

    B CAPEVALSCORES ANDPER-CAPTIONERDOWNSTREAMSCORES B.1 CAPEVAL SCORES Table 3 reports Coverage and Precision scores for all 14 captioners across four semantic domains. 19 Table 4: Normalized understanding and generation benchmark scores. The CAPEval scores (Cand P), aggregate understanding score (U), and aggregate generation score (G) are highlighted in gra...

    2048

  2. [5]

    Sharegpt4v: Improving large multi-modal models with better captions

    10 Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. InEuropean Conference on Com- puter Vision, pp. 370–387. Springer, 2024a. Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin...

    Pith/arXiv arXiv

  3. [7]

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

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261,

    Pith/arXiv arXiv

  4. [8]

    Gemini 3.1 pro model card

    Google DeepMind. Gemini 3.1 pro model card. Technical report, Google DeepMind, 2026a. URL https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/ gemini/3-1-pro. Accessed: 2026-07-26. Google DeepMind. Gemini 3.5 flash model card. Technical report, Google DeepMind, 2026b. URL https://docs.cloud.google.com/gemini-enterprise-agent-platform/mod...

    Pith/arXiv arXiv 2026

  5. [9]

    Flux-reason-6m & prism-bench: A million-scale text-to-image reasoning dataset and comprehensive benchmark.arXiv preprint arXiv:2509.09680,

    Rongyao Fang, Aldrich Yu, Chengqi Duan, Linjiang Huang, Shuai Bai, Yuxuan Cai, Kun Wang, Si Liu, Xihui Liu, and Hongsheng Li. Flux-reason-6m & prism-bench: A million-scale text-to-image reasoning dataset and comprehensive benchmark.arXiv preprint arXiv:2509.09680,

    Pith/arXiv arXiv

  6. [10]

    Datacomp-vlm: Improved open datasets for vision-language models.arXiv preprint arXiv:2606.28551,

    Matteo Farina, Vishaal Udandarao, Thao Nguyen, Selim Kuzucu, Maximilian B ¨other, Andreas Hochlehnert, Adhiraj Ghosh, Marianna Nezhurina, Karsten Roth, Joschka Struber, et al. Datacomp-vlm: Improved open datasets for vision-language models.arXiv preprint arXiv:2606.28551,

    Pith/arXiv arXiv

  7. [11]

    Gavel: Grounded caption error verification and localiza- tion.arXiv preprint arXiv:2606.26923,

    Zixian Gao, Atsushi Hashimoto, and Kuniaki Saito. Gavel: Grounded caption error verification and localiza- tion.arXiv preprint arXiv:2606.26923,

    Pith/arXiv arXiv

  8. [12]

    Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models.arXiv preprint arXiv:2310.14566,

    Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models.arXiv preprint arXiv:2310.14566,

    Pith/arXiv arXiv

  9. [13]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. InProceedings of the 2021 conference on empirical methods in natural language processing, pp. 7514–7528,

    2021

  10. [15]

    Ella: Equip diffusion models with llm for enhanced semantic alignment.arXiv preprint arXiv:2403.05135,

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment.arXiv preprint arXiv:2403.05135,

    Pith/arXiv arXiv

  11. [16]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report.arXiv preprint arXiv:2409.12186,

    Pith/arXiv arXiv

  12. [17]

    Toward robust hyper-detailed image captioning: A multiagent approach and dual evaluation metrics for factuality and coverage.arXiv preprint arXiv:2412.15484,

    Saehyung Lee, Seunghyun Yoon, Trung Bui, Jing Shi, and Sungroh Yoon. Toward robust hyper-detailed image captioning: A multiagent approach and dual evaluation metrics for factuality and coverage.arXiv preprint arXiv:2412.15484,

    Pith/arXiv arXiv

  13. [18]

    Seed-bench: Benchmarking multimodal llms with generative comprehension.arXiv preprint arXiv:2307.16125, 2023a

    Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension.arXiv preprint arXiv:2307.16125, 2023a. Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational...

    Pith/arXiv arXiv

  14. [20]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 26296–26306, 2024a. Jinyu Liu, Xincheng Shuai, Henghui Ding, and Yu-Gang Jiang. Unison: Benchmarking unified multimodal models via synergistic understanding and g...

    Pith/arXiv arXiv

  15. [21]

    Accessed: 2026-07-26

    URLhttps://openai.com/index/ gpt-5-5-system-card/. Accessed: 2026-07-26. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pp. 311–318,

    2026

  16. [24]

    Cambrian-1: A fully open, vision- centric exploration of multimodal llms

    Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, Adithya Jairam Iyer, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision- centric exploration of multimodal llms. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024a. Shengbang Tong, Zhuang Liu, Yue...

    Pith/arXiv arXiv

  17. [25]

    Grasp any region: Towards precise, contextual pixel understanding for multimodal llms.arXiv preprint arXiv:2510.18876, 2025a

    Haochen Wang, Yuhao Wang, Tao Zhang, Yikang Zhou, Yanwei Li, Jiacong Wang, Jiani Zheng, Ye Tian, Jiahao Meng, Zilong Huang, et al. Grasp any region: Towards precise, contextual pixel understanding for multimodal llms.arXiv preprint arXiv:2510.18876, 2025a. 13 Haochen Wang, Yucheng Zhao, Tiancai Wang, Haoqiang Fan, Xiangyu Zhang, and Zhaoxiang Zhang. Ross3...

    arXiv 2025

  18. [26]

    Vgr: Visual grounded reasoning.arXiv preprint arXiv:2506.11991, 2025d

    Jiacong Wang, Zijian Kang, Haochen Wang, Haiyong Jiang, Jiawen Li, Bohong Wu, Ya Wang, Jiao Ran, Xiao Liang, Chao Feng, et al. Vgr: Visual grounded reasoning.arXiv preprint arXiv:2506.11991, 2025d. Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Jiaqi Wang, Haiyang Xu, Ming Yan, Ji Zhang, et al. Amber: An llm-free multi-dimensional...

    Pith/arXiv arXiv

  19. [27]

    Internvl3

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025e. Yana Wei, Hongbo Peng, Yanlin Lai, Liang Zhao, Kangheng Lin, En Yu, Keyu Lv, Han Zhou, Y...

    Pith/arXiv arXiv

  20. [28]

    Qwen-image technical report.arXiv preprint arXiv:2508.02324,

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324,

    Pith/arXiv arXiv

  21. [29]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025a

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025a. Penghui Yang, Long Xing, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Yibin Wang, Yujie Zhou, Jiazi Bu, Jianze Liang, Qidong Huang, et al. Caprl++: Unified reinforcement learn...

    Pith/arXiv arXiv

  22. [30]

    Captionqa: Is your caption as useful as the image itself?arXiv preprint arXiv:2511.21025, 2025b

    Shijia Yang, Yunong Liu, Bohan Zhai, Ximeng Sun, Zicheng Liu, Emad Barsoum, Manling Li, and Chenfeng Xu. Captionqa: Is your caption as useful as the image itself?arXiv preprint arXiv:2511.21025, 2025b. Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference o...

    Pith/arXiv arXiv

  23. [31]

    Actor as its own critic: Unifying region understanding and localization via cyclegrpo.arXiv preprint arXiv:2607.11581,

    Xin Zhang, Haochen Wang, Yikang Zhou, Jason Li, and Robby T Tan. Actor as its own critic: Unifying region understanding and localization via cyclegrpo.arXiv preprint arXiv:2607.11581,

    Pith/arXiv arXiv

  24. [32]

    Low-hallucination synthetic captions for large-scale vision-language model pre-training.arXiv preprint arXiv:2504.13123,

    Xinsong Zhang, Yarong Zeng, Xinting Huang, Hu Hu, Runquan Xie, Han Hu, and Zhanhui Kang. Low-hallucination synthetic captions for large-scale vision-language model pre-training.arXiv preprint arXiv:2504.13123,

    Pith/arXiv arXiv

  25. [2002]

    Aloha: A new measure for hallucination in captioning models

    Suzanne Petryk, David Chan, Anish Kachinthaya, Haodi Zou, John Canny, Joseph Gonzalez, and Trevor Dar- rell. Aloha: A new measure for hallucination in captioning models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies (Volume 2: Short Papers), pp. 342–357,

    2024

  26. [2013]

    Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006,

    Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, et al. Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006,

    Pith/arXiv arXiv

  27. [2015]

    Caparena: Benchmarking and analyzing detailed image captioning in the llm era

    Kanzhi Cheng, Wenpo Song, Jiaxin Fan, Zheng Ma, Qiushi Sun, Fangzhi Xu, Chenyang Yan, Nuo Chen, Jianbing Zhang, and Jiajun Chen. Caparena: Benchmarking and analyzing detailed image captioning in the llm era. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 14077–14094,

    2025

  28. [2020]

    Longcat-image technical report.arXiv preprint arXiv:2512.07584,

    Meituan LongCat Team, Hanghang Ma, Haoxian Tan, Jiale Huang, Junqiang Wu, Jun-Yan He, Lishuai Gao, Songlin Xiao, Xiaoming Wei, Xiaoqi Ma, et al. Longcat-image technical report.arXiv preprint arXiv:2512.07584,

    Pith/arXiv arXiv

  29. [2022]

    Llava-onevision-1.5: Fully open framework for democratized multimodal training.arXiv preprint arXiv:2509.23661,

    Xiang An, Yin Xie, Kaicheng Yang, Wenkang Zhang, Xiuwei Zhao, Zheng Cheng, Yirui Wang, Songcen Xu, Changrui Chen, Didi Zhu, et al. Llava-onevision-1.5: Fully open framework for democratized multimodal training.arXiv preprint arXiv:2509.23661,

    Pith/arXiv arXiv

  30. [2023]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

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

    Pith/arXiv arXiv

  31. [2024]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198,

    Pith/arXiv arXiv

  32. [2025]

    Qwen technical report.arXiv preprint arXiv:2309.16609,

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

    Pith/arXiv arXiv

  33. [2026]

    Evaluating object hallucination in large vision-language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. InProceedings of the 2023 conference on empirical methods in natural language processing, pp. 292–305, 2023c. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C Law...

    2023