arxiv: 2604.26614 · v1 · submitted 2026-04-29 · 💻 cs.CV

Recognition: unknown

State Beyond Appearance: Diagnosing and Improving State Consistency in Dial-Based Measurement Reading

Yuanze Hu , Gen Li , Yuqin Lan , Qingchen Yu , Zhichao Yang , Junwei Jing , Zhaoxin Fan , Xiaotie Deng

Authors on Pith no claims yet

Pith reviewed 2026-05-07 11:41 UTC · model grok-4.3

classification 💻 cs.CV

keywords dial readingmultimodal large language modelsstate consistencyrepresentation alignmentgauge readingmeasurement robustnessappearance invariance

0 comments

The pith

Multimodal models read dials inconsistently because they track visual appearance instead of the underlying state value.

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

The paper shows that current multimodal large language models achieve low accuracy on tasks that require reading measurements from dials such as clocks and gauges. Performance collapses further when the camera angle or lighting changes, even though the actual dial position or reading stays fixed. Feature-space analysis finds that the models fail to group identical states together and do not respect the ordering implied by continuous dial values. The authors introduce TriSCA, a three-part alignment method that forces the model to organize its representations by state distance, to use metadata as direct supervision, and to apply objectives that respect state consistency. A reader would care because many practical monitoring and inspection tasks depend on accurate, stable dial interpretation under real-world viewing conditions.

Core claim

Existing MLLMs not only achieve unsatisfactory accuracy on dial-based readout but suffer sharp performance drops under viewpoint and illumination changes even when the underlying dial state remains fixed. Probing reveals that same-state samples under appearance variation are not consistently clustered and neighboring states fail to preserve local structure implied by continuous dial values. The paper therefore proposes TriSCA, which applies state-distance-aware representation alignment, metadata-grounded observation-to-state supervision, and state-aware objective alignment, and demonstrates gains on controlled clock and gauge benchmarks plus an external real-world benchmark.

What carries the argument

TriSCA, the tri-level state-consistent alignment framework that enforces representation organization by intrinsic dial state rather than by visual appearance.

If this is right

Accuracy on dial readout remains stable when viewpoint or illumination changes while the physical state is unchanged.
Internal representations group samples by true measurement value rather than by superficial visual features.
Each of the three alignment components contributes measurably to the observed gains according to ablation results.
The method transfers to an external real-world dial benchmark beyond the controlled training data.

Where Pith is reading between the lines

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

Similar consistency constraints could be applied to other continuous estimation problems such as reading thermometers or analog meters.
The diagnosis points to a general shortcut bias in MLLMs that may affect any task where geometry or invariants matter more than appearance.
One could test whether the same tri-level alignment improves robustness in video sequences where dial states evolve over time.

Load-bearing premise

That controlled clock and gauge benchmarks together with feature-space probing accurately reflect how the models will behave on uncontrolled real-world dial images.

What would settle it

Collect a new test set of dial images that vary in background, angle, and lighting while keeping the true reading fixed; if models trained with TriSCA still show accuracy drops and poor state clustering comparable to baseline models, the claim of improved state consistency would be refuted.

Figures

Figures reproduced from arXiv: 2604.26614 by Gen Li, Junwei Jing, Qingchen Yu, Xiaotie Deng, Yuanze Hu, Yuqin Lan, Zhaoxin Fan, Zhichao Yang.

**Figure 1.** Figure 1: Current MLLMs are fragile under appearance shifts. view at source ↗

**Figure 2.** Figure 2: Dial-based readout is structured by state relations, view at source ↗

**Figure 3.** Figure 3: Controlled synthesis enables state-aware diagnosis and training for both clocks and gauges. (a) Our synthesis pipeline view at source ↗

**Figure 4.** Figure 4: State-aware alignment reshapes the feature space toward dial-state consistency. (a) Off-the-shelf visual features are view at source ↗

**Figure 5.** Figure 5: State-consistent alignment slows performance view at source ↗

**Figure 6.** Figure 6: Representative case studies showing improved state consistency and local precision after alignment. view at source ↗

read the original abstract

Multimodal large language models (MLLMs) have achieved impressive progress on general multimodal tasks, yet they remain brittle on dial-based measurement reading. In this paper, we study this problem through controlled benchmarks and feature-space probing, and show that current MLLMs not only achieve unsatisfactory accuracy on dial-based readout, but also suffer sharp performance drops under viewpoint and illumination changes even when the underlying dial state remains fixed. Our probing analysis further reveals that same-state samples under appearance variation are not consistently clustered, while neighboring states fail to preserve the local structure implied by continuous dial values. These findings suggest that existing MLLMs largely ignore the intrinsic state geometry of dial measurement tasks and instead rely on superficial appearance cues. Motivated by this diagnosis, we propose TriSCA, a tri-level state-consistent alignment framework for dial-based measurement reading. Specifically, TriSCA consists of state-distance-aware representation alignment, metadata-grounded observation-to-state supervision, and state-aware objective alignment. Extensive ablation studies and evaluation experiments on controlled clock and gauge benchmarks, together with evaluation on an external real-world benchmark, demonstrate the effectiveness of our method.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper diagnoses MLLMs ignoring dial state geometry in favor of appearance and offers TriSCA as a practical alignment fix, but the probing evidence is only suggestive.

read the letter

This paper shows that MLLMs are unreliable for reading measurements from dials because they key off visual appearance instead of the actual state of the dial, and it introduces a three-part alignment method called TriSCA to push the models toward better state consistency. The new part is the diagnosis through controlled tests and feature probing on clock and gauge images, where they demonstrate performance collapse under viewpoint or lighting shifts with fixed states, and inconsistent clustering in feature space for same states. TriSCA then adds state-distance-aware alignment, supervision from metadata, and objective alignment to fix it. They run ablations and test on both synthetic and real benchmarks, which is solid for an applied paper. It does well at highlighting a practical robustness gap that matters for automation and monitoring applications. The experiments are structured with ablations that let you see what helps, and including a real-world benchmark is a plus. The main soft spot is whether the probing truly shows ignorance of state geometry rather than just non-robust encoding of it. The accuracy drops are clear, but the clustering results could still reflect some geometry that's entangled with appearance rather than complete disregard for it. Without tighter controls that hold low-level statistics fixed while varying only dial angle, the diagnosis stays suggestive. The gains from TriSCA look real on the reported benchmarks. This work is aimed at researchers focused on making MLLMs reliable for specific vision tasks like instrument reading. A reader interested in robustness or domain-specific alignment would get value from the benchmarks and method details. It deserves serious referee time because it tackles a concrete problem with experiments and a clear proposal, even if the interpretation of the probing needs tightening. I would send it out for peer review.

Referee Report

2 major / 2 minor

Summary. The paper claims that current multimodal large language models (MLLMs) are brittle on dial-based measurement reading because they ignore intrinsic state geometry and rely on superficial appearance cues. This is diagnosed via controlled clock/gauge benchmarks showing accuracy drops under viewpoint/illumination shifts with fixed states, plus feature-space probing revealing inconsistent clustering of same-state samples and broken local continuity for neighboring states. Motivated by the diagnosis, the authors introduce TriSCA, a tri-level state-consistent alignment framework with state-distance-aware representation alignment, metadata-grounded observation-to-state supervision, and state-aware objective alignment. Ablations and evaluations on the controlled benchmarks plus an external real-world benchmark are presented to demonstrate effectiveness.

Significance. If the central diagnosis and gains hold, the work would be significant for improving MLLM robustness on geometry-sensitive measurement tasks common in industrial, medical, and scientific applications. The combination of controlled benchmarks, feature probing for diagnosis, and a concrete tri-level alignment method with ablations and external validation is a strength; it provides both insight into failure modes and a practical mitigation strategy. The approach could influence future work on state-aware multimodal representations beyond dials.

major comments (2)

[Probing analysis section] Probing analysis (feature-space clustering results): the claim that models ignore state geometry rather than encoding it in an appearance-entangled subspace rests on the assumption that the probed features (unspecified layer/pooling in the vision encoder) isolate internal state representations. Without layer-wise ablations or controls that vary only dial angle while holding low-level appearance statistics fixed, the evidence cannot distinguish the two, directly affecting the motivation for TriSCA's three alignment terms.
[Experiments and benchmarks section] Controlled benchmarks and evaluation: the reported accuracy drops under viewpoint/illumination changes and the post-TriSCA gains are load-bearing for the central claim, yet the manuscript does not detail data splits, exact generation of appearance variations, or whether the same-state samples were generated with matched low-level statistics. This leaves open the possibility that benchmark design influences the diagnosis and claimed improvements.

minor comments (2)

[Method section] The three components of TriSCA are described at a high level in the abstract and method; explicit equations or pseudocode for the loss terms (especially state-distance-aware alignment and state-aware objective) would improve reproducibility.
[Figures] Figure captions for probing visualizations should specify the dimensionality reduction technique, distance metric, and which encoder layer is used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of experimental rigor and probing analysis that we have addressed through revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Probing analysis section] Probing analysis (feature-space clustering results): the claim that models ignore state geometry rather than encoding it in an appearance-entangled subspace rests on the assumption that the probed features (unspecified layer/pooling in the vision encoder) isolate internal state representations. Without layer-wise ablations or controls that vary only dial angle while holding low-level appearance statistics fixed, the evidence cannot distinguish the two, directly affecting the motivation for TriSCA's three alignment terms.

Authors: We acknowledge that the original manuscript did not explicitly specify the probed layer and pooling operation. In the revised version, we have added: (i) explicit details that features are taken from the final layer of the vision encoder with global average pooling; (ii) new layer-wise probing ablations across intermediate and final layers, confirming that inconsistent clustering of same-state samples and broken local continuity persist regardless of layer; (iii) expanded description of the benchmark synthesis process (fixed dial angle/state with independent randomization of viewpoint, illumination, and low-level appearance statistics such as texture and lighting). These controls ensure appearance variations are decoupled from state, supporting that the observed inconsistencies reflect a lack of robust state geometry encoding rather than mere entanglement in a subspace. This bolsters the motivation for TriSCA. revision: yes
Referee: [Experiments and benchmarks section] Controlled benchmarks and evaluation: the reported accuracy drops under viewpoint/illumination changes and the post-TriSCA gains are load-bearing for the central claim, yet the manuscript does not detail data splits, exact generation of appearance variations, or whether the same-state samples were generated with matched low-level statistics. This leaves open the possibility that benchmark design influences the diagnosis and claimed improvements.

Authors: We agree that the original manuscript lacked sufficient detail on benchmark construction, which could raise questions about potential confounds. The revised manuscript now includes a dedicated subsection with: (i) full specification of the synthetic data generation pipeline (parameterized camera poses, lighting models, and texture randomization in Blender while holding dial state fixed); (ii) exact train/validation/test splits (70/15/15, ensuring no state or appearance leakage); (iii) explicit confirmation and examples that same-state samples are generated with deliberately unmatched low-level statistics (varied backgrounds, shadows, and color distributions). We have also made the benchmark generation code publicly available. These additions demonstrate that the accuracy drops and TriSCA gains arise from state inconsistency under appearance shifts, not from benchmark artifacts. revision: yes

Circularity Check

0 steps flagged

No circularity: diagnosis and method are independent of each other

full rationale

The paper's chain proceeds from controlled benchmarks and feature-space probing (accuracy drops under appearance shifts; clustering failures for same-state samples) to a diagnosis that models ignore state geometry, then to proposing TriSCA's three alignment terms. No equations, fitted parameters, or self-citations are shown that reduce the claimed improvement or the diagnosis to a quantity defined by the same inputs. The probing observations and the subsequent alignment framework remain distinct; the former supplies empirical motivation while the latter supplies an independent architectural response. This satisfies the default expectation of a self-contained derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so concrete free parameters, axioms, or invented entities cannot be extracted. The framework description implies tunable alignment weights and loss coefficients but none are named or justified here.

pith-pipeline@v0.9.0 · 5517 in / 1065 out tokens · 50990 ms · 2026-05-07T11:41:52.139647+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

46 extracted references · 26 canonical work pages · 10 internal anchors

[1]

Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al . 2025. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631(2025)

work page internal anchor Pith review arXiv 2025
[2]

Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. 2023. Shikra: Unleashing Multimodal LLM’s Referential Dialogue Magic. arXiv:2306.15195 [cs.CV] https://arxiv.org/abs/2306.15195 State Beyond Appearance: Diagnosing and Improving State Consistency in Dial-Based Measurement Reading arXiv, 2026, Table 4: Continuous-state evaluation on...

work page internal anchor Pith review arXiv 2023
[3]

Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. 2023. ShareGPT4V: Improving Large Multi-Modal Models with Better Captions. arXiv:2311.12793 [cs.CV] https://arxiv.org/abs/2311.12793

work page internal anchor Pith review arXiv 2023
[4]

Zihui Cheng, Qiguang Chen, Xiao Xu, Jiaqi Wang, Weiyun Wang, Hao Fei, Yidong Wang, Alex Jinpeng Wang, Zhi Chen, Wanxiang Che, et al. 2025. Visual thoughts: A unified perspective of understanding multimodal chain-of-thought.arXiv preprint arXiv:2505.15510(2025)

work page arXiv 2025
[5]

Zixu Cheng, Jian Hu, Ziquan Liu, Chenyang Si, Wei Li, and Shaogang Gong. 2025. V-star: Benchmarking video-llms on video spatio-temporal reasoning.arXiv preprint arXiv:2503.11495(2025)

work page arXiv 2025
[6]

Gonzalez, Ion Stoica, and Eric P

Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality. https://lmsys.org/blog/2023-03-30-vicuna/

2023
[7]

Jaeha Choi, Jin Won Lee, Siwoo You, and Jangho Lee. 2026. It’s Time to Get It Right: Improving Analog Clock Reading and Clock-Hand Spatial Reasoning in Vision-Language Models.arXiv preprint arXiv:2603.08011(2026)

work page arXiv 2026
[8]

Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. 2023. DoLa: Decoding by Contrasting Layers Improves Fac- tuality in Large Language Models. arXiv:2309.03883 [cs.CL] https://arxiv.org/ abs/2309.03883

work page arXiv 2023
[9]

Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multi- modality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261(2025)

work page internal anchor Pith review arXiv 2025
[10]

Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al
[11]

InProceedings of the Computer Vision and Pattern Recognition Conference

Molmo and pixmo: Open weights and open data for state-of-the-art vision- language models. InProceedings of the Computer Vision and Pattern Recognition Conference. 91–104
[12]

Tairan Fu, Miguel González, Javier Conde, Elena Merino-Gómez, and Pedro Reviriego. 2025. Have Multimodal Large Language Models (MLLMs) Really Learned to Tell the Time on Analog Clocks?IEEE Internet Computing(2025)

2025
[13]

Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. 2024. Blink: Multimodal large language models can see but not perceive. InEuropean Conference on Computer Vision. Springer, 148–166

2024
[14]

Anisha Gunjal, Jihan Yin, and Erhan Bas. 2023. Detecting and Preventing Hallucinations in Large Vision Language Models. arXiv:2308.06394 [cs.CV] https://arxiv.org/abs/2308.06394

work page arXiv 2023
[15]

Hulingxiao He, Zijun Geng, and Yuxin Peng. [n. d.]. Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning. InThe Fourteenth International Conference on Learning Representations
[16]

Yuanze Hu, Zhaoxin Fan, Xinyu Wang, Gen Li, Ye Qiu, Zhichao Yang, Wenjun Wu, Kejian Wu, Yifan Sun, Xiaotie Deng, et al. 2025. TinyAlign: Boosting Lightweight Vision-Language Models by Mitigating Modal Alignment Bottlenecks.arXiv preprint arXiv:2505.12884(2025)

work page arXiv 2025
[17]

Ryo Kamoi, Yusen Zhang, Sarkar Snigdha Sarathi Das, Ranran Haoran Zhang, and Rui Zhang. 2024. Visonlyqa: Large vision language models still struggle with visual perception of geometric information.arXiv preprint arXiv:2412.00947 (2024)

work page arXiv 2024
[18]

Aditya Sanjiv Kanade and Tanuja Ganu. 2026. Do you see me: A multidimensional benchmark for evaluating visual perception in multimodal llms. InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 7285–7326

2026
[19]

Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning.Advances in neural information processing systems33 (2020), 18661– 18673

2020
[20]

Florian Leiser, Sven Eckhardt, Valentin Leuthe, Merlin Knaeble, Alexander Maed- che, Gerhard Schwabe, and Ali Sunyaev. 2024. HILL: A Hallucination Identifier for Large Language Models. arXiv:2403.06710 [cs.HC] https://arxiv.org/abs/2403. 06710

work page arXiv 2024
[21]

Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. 2023. Mitigating Object Hallucinations in Large Vision- Language Models through Visual Contrastive Decoding. arXiv:2311.16922 [cs.CV] https://arxiv.org/abs/2311.16922

work page arXiv 2023
[22]

Juan Leon-Alcazar, Yazeed Alnumay, Cheng Zheng, Hassane Trigui, Sahejad Patel, and Bernard Ghanem. 2024. Learning to read analog gauges from synthetic data. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 8616–8625

2024
[23]

Bin Lin, Zhenyu Tang, Yang Ye, Jinfa Huang, Junwu Zhang, Yatian Pang, Peng Jin, Munan Ning, Jiebo Luo, and Li Yuan. 2024. MoE-LLaVA: Mixture of Experts for Large Vision-Language Models. arXiv:2401.15947 [cs.CV] https://arxiv.org/ abs/2401.15947

work page arXiv 2024
[24]

Fenfen Lin, Yesheng Liu, Haiyu Xu, Chen Yue, Zheqi He, Mingxuan Zhao, Miguel Hu Chen, Jiakang Liu, JG Yao, and Xi Yang. 2025. Do Vision-Language Models Measure Up? Benchmarking Visual Measurement Reading with Mea- sureBench.arXiv preprint arXiv:2510.26865(2025)

work page arXiv 2025
[25]

Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang
[26]

Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning

Mitigating Hallucination in Large Multi-Modal Models via Robust Instruc- tion Tuning. arXiv:2306.14565 [cs.CV] https://arxiv.org/abs/2306.14565

work page internal anchor Pith review arXiv
[27]

Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. 2024. Ocrbench: on the hidden mystery of ocr in large multimodal models.Science China Information Sciences67, 12 (2024), 220102

2024
[28]

Maurits Reitsma, Julian Keller, Kenneth Blomqvist, and Roland Siegwart. 2024. Under pressure: learning-based analog gauge reading in the wild. In2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 14–20

2024
[29]

Gabriel Salomon, Rayson Laroca, and David Menotti. 2020. Deep learning for image-based automatic dial meter reading: Dataset and baselines. In2020 Inter- national joint conference on neural networks (IJCNN). IEEE, 1–8

2020
[30]

Gabriel Salomon, Rayson Laroca, and David Menotti. 2022. Image-based auto- matic dial meter reading in unconstrained scenarios.Measurement204 (2022), 112025

2022
[31]

Rohit Saxena, Aryo Pradipta Gema, and Pasquale Minervini. 2025. Lost in time: Clock and calendar understanding challenges in multimodal LLMs.arXiv preprint arXiv:2502.05092(2025)

work page arXiv 2025
[32]

Florian Schroff, Dmitry Kalenichenko, and James Philbin. 2015. Facenet: A unified embedding for face recognition and clustering. InProceedings of the IEEE conference on computer vision and pattern recognition. 815–823

2015
[33]

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov
[34]

Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347 (2017)

work page internal anchor Pith review arXiv 2017
[35]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv, 2026, Hu et al. arXiv:2402.03300(2024)

work page internal anchor Pith review arXiv 2024
[36]

Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. 2019. Towards vqa models that can read. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8317–8326

2019
[37]

Ilias Stogiannidis, Steven McDonagh, and Sotirios A Tsaftaris. 2025. Mind the gap: Benchmarking spatial reasoning in vision-language models.arXiv preprint arXiv:2503.19707(2025)

work page arXiv 2025
[38]

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)

work page internal anchor Pith review arXiv 2024
[39]

Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. 2023. CogVLM: Visual Expert for Pretrained Language Models. arXiv:2311.03079 [cs.CV] https://arxiv.org/abs/2311.03079

work page arXiv 2023
[40]

Yujun Wang, Aniri, Jinhe Bi, Soeren Pirk, and Yunpu Ma. 2025. ASCD: Attention-Steerable Contrastive Decoding for Reducing Hallucination in MLLM. arXiv:2506.14766 [cs.CV] https://arxiv.org/abs/2506.14766

work page arXiv 2025
[41]

Charig Yang, Weidi Xie, and Andrew Zisserman. 2022. It’s about time: Analog clock reading in the wild. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2508–2517

2022
[42]

Zhichao Yang, Zhaoxin Fan, Gen Li, Yuanze Hu, Xinyu Wang, Ye Qiu, Xin Wang, Yifan Sun, and Wenjun Wu. 2025. Can Structured Templates Facilitate LLMs in Tackling Harder Tasks?: An Exploration of Scaling Laws by Difficulty.arXiv preprint arXiv:2508.19069(2025)

work page arXiv 2025
[43]

Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. 2024. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9556–9567

2024
[44]

Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. 2025. MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Wa...

2025
[45]

Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. 2023. Multimodal chain-of-thought reasoning in language models.arXiv preprint arXiv:2302.00923(2023)

work page internal anchor Pith review arXiv 2023
[46]

Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models. arXiv:2304.10592 [cs.CV] https://arxiv.org/abs/2304.10592

work page internal anchor Pith review arXiv 2023