Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Frontier vision-language models refuse to believe a perfect detector when an essential object part is missing.

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 14:40 UTC pith:UF4RLLBD

load-bearing objection Useful existence proof that SOTA VLMs miss removed essential parts; the tool-evidence 'cannot be corrected' claim overreaches because the simulated detector message is weak evidence, not a strong contradiction. the 4 major comments →

arxiv 2607.18673 v2 pith:UF4RLLBD submitted 2026-07-21 cs.CV

MissingBench-Verified: Probing Vision-Language Models' Inability to Detect Missing Object Parts

Wenqi Marshall Guo , Qingyun Qian , Shiyu Zhou , Guoping Luo , Shan Du This is my paper
classification cs.CV
keywords missing object partsvisual hallucinationvision-language modelsMissingBench-Verifiedexistence biastool-assisted verificationparametric knowledgemultimodal robustness
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.

This paper argues that leading vision-language models systematically fail to notice when an essential part of an object has been removed—even when an external tool explicitly tells them the part is absent. It introduces a small human-verified benchmark, MissingBench-Verified, of 118 edited real images where an essential component is gone. Across ten frontier models, none identified the absence more than 75% of the time, and several said the removed part was 'clearly visible' more often than they said it was missing. Injecting a simulated perfect detection result, providing image-editing tools, increasing reasoning effort, or fine-tuning on easier hallucination data all produced negligible improvement. The paper concludes that the failure is rooted in internal prior knowledge and visual processing, requiring architectural or training-level changes rather than prompting or post-hoc correction.

Core claim

The central discovery is that current frontier VLMs exhibit a strong existence bias for essential object parts: after a part like an airplane's engine or a phone's home button is edited out, the models often answer that it is 'clearly visible,' and they keep that answer even when a simulated perfect detector reports 'undetected, low confidence' in the context. Across ten leading models, no model identified the absence more than 75% of the time, and the worst model was correct only 44% of the time. Pre-answer and post-answer tool injection, image cropping and contrast tools, longer reasoning, and fine-tuning on an easier hallucination dataset all failed to raise performance meaningfully. The

What carries the argument

MissingBench-Verified—a human-verified set of 118 real photos in which an essential component has been digitally or manually removed—is the test bed. The evaluation protocol asks each model a three-choice visibility question and forces it to produce reasoning in JSON. The key mechanism that isolates the failure is the simulated perfect detector: a text line saying 'undetected, low confidence' is injected into the context either before or after the model's first answer. This gives the model a maximally strong external signal that the object is absent, so any persistent 'visible' answer demonstrates that internal priors override external evidence. A third setting lets the model call image-proc

Load-bearing premise

The conclusion that external tool evidence cannot correct the hallucination rests on treating a text-injected simulated perfect detector as equivalent to a real detector's visually grounded output; if models trust real tool outputs more than text messages, the conclusion may not hold.

What would settle it

Run MissingBench-Verified through a VLM-agent loop in which an actual open-vocabulary detector returns a marked-up image (no bounding box, low confidence) as a tool result before the model answers, and check whether 'not visible' accuracy rises above 75%. If it does, the paper's strongest claim is falsified; if it stays below, the claim is supported.

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

If this is right

  • Current VLMs are not reliable for automated visual inspection or monitoring when essential components may be absent; every tested model failed on a substantial share of cases.
  • Tool-augmented pipelines inherit the model's bias: even a simulated perfect detector leaves most models below 75% accuracy on detecting absence.
  • More reasoning time does not help; accuracy stayed flat or slightly decreased as reasoning effort increased from none to maximum.
  • Fine-tuning on related, easier hallucination data does not transfer to this harder setting.
  • Architectural or training-level changes are needed so models can override internal expectations when visual evidence contradicts them.

Where Pith is reading between the lines

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

  • The tool-conflict result is an upper bound on real tool-augmented systems: the model saw the tool message as text, not as a spatially grounded detection visualization. A real detector returning a marked-up image may carry more perceptual weight, so the conclusion that 'tools cannot help' should be read as 'textual tool reports cannot help.'
  • The benchmark's design implies a testable gradient: the more an object part is implied by world knowledge (airplane engines, phone buttons), the stronger the existence bias. A natural extension is to vary the 'essentialness' of the removed part and measure whether the failure rate tracks it.
  • If this bias is as deep as reported, it suggests that multimodal pretraining rewards statistical co-occurrence of parts and objects so strongly that it outweighs the visual signal. That points to a training-level fix—such as augmenting data with missing-part examples or counterfactual supervision—rather than inference-time intervention.

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 / 4 minor

Summary. MissingBench-Verified evaluates whether state-of-the-art VLMs recognize when an essential object part is absent. The authors construct 118 images (real photos and internet images) from which a human-specified element is removed via an image editor and human-verified. They query ten VLMs (proprietary and open-source) under several settings: direct query, pre-answer injection of a simulated 'perfect detector' result, post-answer injection, and 'thinking with images' tool use; they also vary reasoning effort for GPT-5.2 and compare against OWLv2/YOLOE. The main reported findings are that all ten models answer 'visible' at high rates on edited images; simulated detector injections improve some models but leave most below 75% accuracy; image tools give negligible gains; and object detectors exhibit poor sensitivity. The paper claims this demonstrates a fundamental limitation that cannot be corrected by current prompting or post-hoc correction techniques.

Significance. If the core existence result holds, the benchmark is a valuable stress test: it targets a realistic inspection scenario, uses real images rather than synthetic degradation, includes human verification, compares ten frontier models, and anchors difficulty against RH-Bench. The finding that leading VLMs frequently report a missing essential part as present, even after cropping or enhancement, is important for deployment in monitoring and inspection tasks. However, the broader conclusions—that external tool evidence 'explicitly contradicts' model perception and that the failure mode is uncorrectable—depend on an unvalidated weak-evidence simulation and are not supported as currently stated. The work is creditworthy for its dataset construction and multi-model evaluation, but the claims need to be narrowed and statistical support added.

major comments (4)
  1. [§2.2, Tables 3–4] The simulated detector is described as 'representing a perfect detector,' but the injected message is {'detection confidence': 0.03, 'result': '(<0.1) undetected'}. A 'not detected above 0.1' result does not assert absence; it asserts only that no confident detection was produced. A model that answers 'clearly visible' while acknowledging the low-confidence tool result (Figure 1) is not overriding contradictory evidence; it is discounting weak evidence. Consequently the phrases 'external tool evidence explicitly contradicts the model's visual perception' and the conclusion that tool-assisted verification cannot correct this failure are not supported. At minimum, add a condition with a high-confidence absence message (e.g., confidence 0.99, result: absent) or real detector output; the current design is not an upper bound for tool-augmented systems.
  2. [§2.1, §3.1] The dataset has 118 edited images and all headline rates are reported as percentages without confidence intervals, error bars, or significance tests, while the abstract and conclusion call the failure rates 'significant.' With 118 trials, a difference of about 4 percentage points (e.g., ±4.3 in Table 3) is within sampling noise; the text acknowledges some decreases 'likely reflect noise' but still draws strong aggregate claims. Report Wilson intervals or exact tests, and state the human-verification baseline for the edited images (did a human correctly identify the part as absent in every case? at what agreement?). Without this, the 'consistent and significant' claim is not quantitatively supported.
  3. [§3.1, §5] The conclusion that 'fine-tuning on an easier dataset' provides negligible improvement and that the failure 'cannot be addressed through current prompting or post-hoc correction techniques' is too strong. The only fine-tuned model evaluated is NegAwareQwen from VSF, which was trained for a different distribution of missing objects; no model in the paper was fine-tuned on MissingBench-Verified or on a comparable hard-missing-part dataset. A null result on one off-the-shelf fine-tuned model does not license a claim about fine-tuning as a family. Please either remove the fine-tuning claim or test a model actually trained for this task.
  4. [§3.1, Table 7] The framing that a non-significant p-value 'strengthens our point' is statistically invalid: failing to reject the null is not evidence for the null, especially when the point estimate shows a 7.7-point drop between Low and xHigh for the 'not visible' rate. The reasoning-effort result should be reported as inconclusive, or accompanied by an equivalence test or power analysis, rather than as confirmation that the hallucination is 'embedded in the system.'
minor comments (4)
  1. [§3.1] There is a placeholder 'Section??' for the detailed failure analysis of why image processing tools did not help; this section is missing and should be provided or the reference removed.
  2. [Throughout] Typographical issues: 'Owlv2' should be 'OWLv2,' 'thershold' should be 'threshold,' 'Furture Work' should be 'Future Work.'
  3. [Table 6] The notation t0/t1 and 'Original Acc'/'Edited Acc' is confusing. The caption should define each threshold and clarify that the default threshold differs between OWLv2 (0.3) and YOLOE (0.25).
  4. [Table 1] The dash markers in the 'Think with Images' column could be clearer; consider a separate column or footnote explaining that only three models were tested in that setting.

Circularity Check

0 steps flagged

No significant circularity: benchmark inputs, model evaluations, and external comparisons are independent of the reported failure rates.

full rationale

The paper's central result is a measured benchmark outcome: human-curated, human-verified images of objects with removed essential parts are shown to ten externally obtained VLMs, and the models' responses are recorded as 'Visible / Hard to See / Not Visible.' Nothing in the construction forces a particular answer: the ground truth ('Not Visible') is defined by the image-editing procedure, and the models are not trained or fitted on this benchmark. The comparison to RH-Bench is an external anchor, and the object-detection experiments (OWLv2, YOLOE) are independent baselines. The only self-citation, VSF Guo & Du (2026), is used as prior motivation and its NegAwareQwen model is included merely as an additional baseline; the paper actually reports that this self-cited model 'performed very poorly,' which is the opposite of using the self-citation to prop up the claim. The simulated 'perfect detector' injection is a design assumption that may be criticized as weak evidence ('low confidence, undetected' rather than a high-confidence absence assertion), but that is a validity or strength-of-intervention concern, not a circularity: the injected text is not fitted to the models' answers and does not by construction determine the measured NV rates. The incomplete cross-reference 'Section??' is a typographical defect, not a circular step. Therefore the derivation chain is self-contained with respect to the benchmark's empirical claims, and the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The paper's claims rest on dataset curation and on the simulated-detector proxy; both are domain assumptions. There are no fitted parameters in the VLM-evaluation portion, aside from detector thresholds used for the object-detection comparison. No new physical or structural entities are introduced.

free parameters (1)
  • object detection threshold t1 for OWLv2/YOLOE = not reported (optimized on test set; Table 6)
    The t1 columns select a threshold to make original-image accuracy exceed 80%; Amax selects the optimal threshold on the test set. This is fit to the evaluation data, though the paper labels it as an upper bound rather than a claimed parameter of the VLM result.
axioms (2)
  • domain assumption The 118 edited images have correct ground truth: the target part is absent and the edit is visually natural.
    Section 2.1 says images are human-verified to filter failed/unatural edits, but no inter-annotator agreement or human detection baseline is reported; every accuracy number assumes 'Not Visible' is the single correct answer.
  • domain assumption A simulated detector message ('undetected, low confidence') injected into the prompt is a valid proxy for a real perfect object detector and an upper bound for tool-augmented approaches.
    Section 2.2 'Pre-answer tool injection'; the paper does not compare against an actual visual tool output, so the tool-resistance conclusion depends on models weighting this text the same as real tool evidence.

pith-pipeline@v1.3.0-alltime-deepseek · 10078 in / 13592 out tokens · 143602 ms · 2026-08-01T14:40:37.097870+00:00 · methodology

0 comments
read the original abstract

Vision Language Models (VLMs) are well known for hallucinating non-existent objects in images. Objects with missing parts present a unique challenge for VLMs, stemming from both real-world knowledge bias and the scarcity of such images in training data. We present MissingBench-Verified, a benchmark designed to evaluate a specific and practically relevant scenario: when vision-language models fail to recognize that an essential component of an object has been removed. Across ten leading models, we observe consistent and significant failure rates that persist even when external tool evidence explicitly contradicts the model's visual perception. We further ask whether granting models access to image processing tools (e.g., cropping, contrast adjustment) enables autonomous inspection to resolve these failures. We find that existing mitigation strategies, including tool-assisted verification, autonomous visual reasoning, longer reasoning durations, and fine-tuning on an easier dataset, provide negligible improvement, indicating that this failure mode cannot be addressed through current prompting or post-hoc correction techniques. Our findings highlight a fundamental limitation of current VLM for inspection and monitoring tasks and underscore the need for architectural or training-level interventions that enable models to override internal expectations when confronted with contradictory evidence.

Figures

Figures reproduced from arXiv: 2607.18673 by Guoping Luo, Qingyun Qian, Shan Du, Shiyu Zhou, Wenqi Marshall Guo.

Figure 1
Figure 1. Figure 1: Vision language models could have severe [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A word cloud of missing objects of our dataset. We construct the dataset using real-world captured photos and internet images, each collected alongside the name of the element to be removed. These im￾ages are not scraped using automated tools but are rather human-chosen images with high selectivity. Elements are removed using two approaches. The majority of images are processed with Nano Banana 2, a SOTA i… view at source ↗
Figure 3
Figure 3. Figure 3: (Top) A false negative example from YOLOE, where it failed to detect the windows start button [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An interesting failure case arises in the “with image processing tools” setting: the LLM correctly [PITH_FULL_IMAGE:figures/full_fig_p008_4.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

46 extracted references · 7 canonical work pages · 2 internal anchors

  1. [1]

    FirstName LastName , title =

  2. [2]

    FirstName Alpher , title =

  3. [3]

    Journal of Foo , volume = 13, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe , title =. Journal of Foo , volume = 13, number = 1, pages =

  4. [4]

    Journal of Foo , volume = 14, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe and FirstName Gamow , title =. Journal of Foo , volume = 14, number = 1, pages =

  5. [5]

    FirstName Alpher and FirstName Gamow , title =

  6. [6]

    Rohrbach, Anna and Hendricks, Lisa Anne and Burns, Kaylee and Darrell, Trevor and Saenko, Kate , month = mar, year =. Object. doi:10.48550/arXiv.1809.02156 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1809.02156

  7. [7]

    Poison as

    Zhang, Kejia and Tao, Keda and Tang, Jiasheng and Wang, Huan , month = feb, year =. Poison as. doi:10.48550/arXiv.2501.19164 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.19164

  8. [8]

    Liu, Chengzhi and Xu, Zhongxing and Wei, Qingyue and Wu, Juncheng and Zou, James and Wang, Xin Eric and Zhou, Yuyin and Liu, Sheng , month = jun, year =. More. doi:10.48550/arXiv.2505.21523 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.21523

  9. [9]

    Xu, Zhe and Wang, Zhicai and Wu, Junkang and Lu, Jinda and Wang, Xiang , month = nov, year =. Causal-. doi:10.48550/arXiv.2511.10268 , abstract =

    doi:10.48550/arxiv.2511.10268

  10. [10]

    Liu, Hanchao and Xue, Wenyuan and Chen, Yifei and Chen, Dapeng and Zhao, Xiutian and Wang, Ke and Hou, Liping and Li, Rongjun and Peng, Wei , month = may, year =. A. doi:10.48550/arXiv.2402.00253 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.00253

  11. [11]

    Computer

    Ye-Bin, Moon and Hyeon-Woo, Nam and Choi, Wonseok and Oh, Tae-Hyun , editor =. Computer. 2025 , pages =. doi:10.1007/978-3-031-73247-8_14 , abstract =

    work page doi:10.1007/978-3-031-73247-8_14 2025

  12. [12]

    and Zhang, Aidong , month = dec, year =

    Ye, Wenqian and Liu, Bohan and Zheng, Guangtao and Wang, Di and Ma, Yunsheng and Cao, Xu and Lai, Bolin and Rehg, James M. and Zhang, Aidong , month = dec, year =. doi:10.1145/3770854.3785678 , abstract =

    arXiv

  13. [13]

    doi:10.48550/arXiv.2508.10931 , abstract =

    Guo, Wenqi and Du, Shan , month = feb, year =. doi:10.48550/arXiv.2508.10931 , abstract =

    doi:10.48550/arxiv.2508.10931

  14. [14]

    Normalized

    Chen, Dar-Yen and Bandyopadhyay, Hmrishav and Zou, Kai and Song, Yi-Zhe , month = jun, year =. Normalized. doi:10.48550/arXiv.2505.21179 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.21179

  15. [15]

    doi:10.48550/arXiv.2412.02687 , abstract =

    Nguyen, Viet and Nguyen, Anh and Dao, Trung and Nguyen, Khoi and Pham, Cuong and Tran, Toan and Tran, Anh , month = dec, year =. doi:10.48550/arXiv.2412.02687 , abstract =

    work page doi:10.48550/arxiv.2412.02687

  16. [16]

    and Carley, Kathleen M

    Carragher, Peter and Rao, Nikitha and Jha, Abhinand and Raghav, R. and Carley, Kathleen M. , month = jun, year =. doi:10.36190/2025.27 , abstract =

    work page doi:10.36190/2025.27 2025

  17. [17]

    Benchmarking

    Jia, Yifan and Jiang, Kailin and Liang, Yuyang and Ren, Qihan and Xin, Yi and Yang, Rui and Feng, Fenze and Chen, Mingcai and Lu, Hengyang and Wang, Haozhe and Qu, Xiaoye and Liu, Dongrui and Cui, Lizhen and Du, Yuntao , month = may, year =. Benchmarking. doi:10.48550/arXiv.2505.19509 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.19509

  18. [18]

    Guo, Wenqi Marshall and Du, Yiyang and Tworek, Heidi J. S. and Du, Shan , month = oct, year =. Position:. doi:10.48550/arXiv.2509.08833 , abstract =

    work page doi:10.48550/arxiv.2509.08833

  19. [19]

    Liu, Xiaoyuan and Wang, Wenxuan and Yuan, Youliang and Huang, Jen-tse and Liu, Qiuzhi and He, Pinjia and Tu, Zhaopeng , month = may, year =. Insight. doi:10.48550/arXiv.2410.08145 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.08145

  20. [20]

    doi:10.48550/arXiv.2404.10198 , abstract =

    Wu, Kevin and Wu, Eric and Zou, James , month = feb, year =. doi:10.48550/arXiv.2404.10198 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.10198

  21. [21]

    Jin, Zhuoran and Cao, Pengfei and Chen, Yubo and Liu, Kang and Jiang, Xiaojian and Xu, Jiexin and Li, Qiuxia and Zhao, Jun , month = feb, year =. Tug-of-. doi:10.48550/arXiv.2402.14409 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.14409

  22. [22]

    Investigating

    Cheng, Jiali and Pan, Rui and Amiri, Hadi , month = jan, year =. Investigating. doi:10.48550/arXiv.2601.09760 , abstract =

    doi:10.48550/arxiv.2601.09760

  23. [23]

    Qwen3.5:

    Team, Qwen , month = feb, year =. Qwen3.5:

  24. [24]

    2026 , note =

    Kimi. 2026 , note =. doi:10.48550/arXiv.2602.02276 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.02276 2026

  25. [25]

    doi:10.48550/arXiv.2507.01006 , abstract =

    2026 , note =. doi:10.48550/arXiv.2507.01006 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.01006 2026

  26. [26]

    doi:10.48550/arXiv.2512.02556 , abstract =

    2025 , note =. doi:10.48550/arXiv.2512.02556 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2512.02556 2025

  27. [27]

    Thinking with images , url =

  28. [28]

    Yang, Wenhao and Xia, Yu and Huang, Jinlong and Lu, Shiyin and Chen, Qing-Guo and Xu, Zhao and Luo, Weihua and Zhang, Kaifu and Wan, Yuanyu and Zhang, Lijun , month = jan, year =. Deep. doi:10.48550/arXiv.2512.17306 , abstract =

    doi:10.48550/arxiv.2512.17306

  29. [29]

    MASH-VLM: Mitigating Action-Scene Hallucination in Video-LLMs through Disentangled Spatial-Temporal Representations

    Bae, Kyungho and Kim, Jinhyung and Lee, Sihaeng and Lee, Soonyoung and Lee, Gunhee and Choi, Jinwoo , month = mar, year =. doi:10.48550/arXiv.2503.15871 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.15871

  30. [30]

    MedHallTune: An Instruction-Tuning Benchmark for Mitigating Medical Hallucination in Vision-Language Models

    Yan, Qiao and Yuan, Yuchen and Hu, Xiaowei and Wang, Yihan and Xu, Jiaqi and Li, Jinpeng and Fu, Chi-Wing and Heng, Pheng-Ann , month = feb, year =. doi:10.48550/arXiv.2502.20780 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.20780

  31. [31]

    Oh, Hongseok and Hwang, Wonseok , month = jan, year =. Do. doi:10.48550/arXiv.2502.20034 , abstract =

    work page doi:10.48550/arxiv.2502.20034

  32. [32]

    Minderer, Matthias and Gritsenko, Alexey and Houlsby, Neil , month = may, year =. Scaling. doi:10.48550/arXiv.2306.09683 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.09683

  33. [33]

    doi:10.48550/arXiv.2505.01431 , abstract =

    Guo, Wenqi and Shehata, Mohamed and Du, Shan , month = jul, year =. doi:10.48550/arXiv.2505.01431 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.01431

  34. [34]

    Proceedings of the

    Guo, Wenqi and Du, Yiyang and Du, Shan , year =. Proceedings of the

  35. [35]

    Chain-of-

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed and Le, Quoc and Zhou, Denny , month = jan, year =. Chain-of-. doi:10.48550/arXiv.2201.11903 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2201.11903

  36. [36]

    Carion, Nicolas and Gustafson, Laura and Hu, Yuan-Ting and Debnath, Shoubhik and Hu, Ronghang and Coll-Vinent, Didac Suris and Ryali, Chaitanya and Alwala, Kalyan Vasudev and Khedr, Haitham and Huang, Andrew and Lei, Jie and Ma, Tengyu and Guo, Baishan and Kalla, Arpit and Marks, Markus and Greer, Joseph and Wang, Meng and Sun, Peize and Rädle, Roman and ...

  37. [37]

    doi:10.48550/arXiv.2501.18954 , abstract =

    Fu, Shenghao and Yang, Qize and Mo, Qijie and Yan, Junkai and Wei, Xihan and Meng, Jingke and Xie, Xiaohua and Zheng, Wei-Shi , month = jan, year =. doi:10.48550/arXiv.2501.18954 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.18954

  38. [38]

    doi:10.48550/arXiv.2503.07465 , abstract =

    Wang, Ao and Liu, Lihao and Chen, Hui and Lin, Zijia and Han, Jungong and Ding, Guiguang , month = oct, year =. doi:10.48550/arXiv.2503.07465 , abstract =

    doi:10.48550/arxiv.2503.07465

  39. [39]

    Evaluating and

    He, Yixiao and Sun, Haifeng and Ren, Pengfei and Wang, Jingyu and Wang, Huazheng and Qi, Qi and Zhuang, Zirui and Wang, Jing , editor =. Evaluating and. Proceedings of the 2025. 2025 , pages =. doi:10.18653/v1/2025.naacl-long.349 , abstract =

    doi:10.18653/v1/2025.naacl-long.349 2025

  40. [40]

    Improved

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Lee, Yong Jae , month = may, year =. Improved. doi:10.48550/arXiv.2310.03744 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.03744

  41. [41]

    doi:10.48550/arXiv.2503.17352 , abstract =

    Deng, Yihe and Bansal, Hritik and Yin, Fan and Peng, Nanyun and Wang, Wei and Chang, Kai-Wei , month = nov, year =. doi:10.48550/arXiv.2503.17352 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.17352

  42. [42]

    Yang, Yi and He, Xiaoxuan and Pan, Hongkun and Jiang, Xiyan and Deng, Yan and Yang, Xingtao and Lu, Haoyu and Yin, Dacheng and Rao, Fengyun and Zhu, Minfeng and Zhang, Bo and Chen, Wei , month = mar, year =. R1-. doi:10.48550/arXiv.2503.10615 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.10615

  43. [43]

    Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae , month = dec, year =. Visual. doi:10.48550/arXiv.2304.08485 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.08485

  44. [44]

    prunaai/z-image-turbo

  45. [45]

    Team, Z.-Image and Cai, Huanqia and Cao, Sihan and Du, Ruoyi and Gao, Peng and Hoi, Steven and Hou, Zhaohui and Huang, Shijie and Jiang, Dengyang and Jin, Xin and Li, Liangchen and Li, Zhen and Li, Zhong-Yu and Liu, David and Liu, Dongyang and Shi, Junhan and Wu, Qilong and Yu, Feng and Zhang, Chi and Zhang, Shifeng and Zhou, Shilin , month = dec, year =....

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.22699

  46. [46]

    Seeing is

    Deng, Ailin and Chen, Zhirui and Hooi, Bryan , month = apr, year =. Seeing is. doi:10.48550/arXiv.2402.15300 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.15300