A Training-Free, Task-Agnostic Framework for Enhancing MLLM Performance on High-Resolution Images

Hanjung Kim; Heechan Choi; Jaeseong Lee; Seonjoo Kim; Yeeun Choi

REVIEW 4 major objections 4 minor 1 cited by

A Training-Free, Task-Agnostic Framework for Enhancing MLLM Performance on High-Resolution Images

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Ask the model where to look, then crop: 4K grounding accuracy jumps 21 points.

desk verdict A simple, honest coarse-to-fine crop-and-predict paper with real gains on 4K/8K MLLM benchmarks; missing a few details that a revision can fix. read the letter →

arxiv 2507.10202 v1 pith:FR2DA3TE submitted 2025-07-14 cs.CV cs.AI

Jaeseong Lee , Yeeun Choi , Heechan Choi , Hanjung Kim , Seonjoo Kim This is my paper

classification cs.CVcs.AI

keywords multimodallargelanguagemodelshigh-resolutionimagestraining-freecoarse-to-fineGUIgroundingvisualperceptioncandidateregionextractiondownsampling

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

The paper proposes a training-free, task-agnostic two-stage framework, Extract Candidate then Predict (ECP), for improving multimodal large language model (MLLM) performance on high-resolution images. The key claim is that even though MLLMs struggle on high-resolution inputs, their predictions on downsampled images contain implicit localization cues that can guide a zoomed-in crop. By first asking the model for a point or bounding box on a downsampled image and then making the final prediction from a 1024x1024 crop centered on that candidate, ECP improves 4K GUI grounding by 21.3 percentage points, 4K perception by 5.8 points, and 8K perception by 5.2 points over single-stage baselines. This matters because it offers a simple, parameter-free way to recover fine-grained accuracy without retraining or task-specific heuristics.

What carries the argument

The mechanism is a two-stage pipeline: Extract Candidate (EC) feeds a downsampled high-resolution image and text instruction to an MLLM, which outputs either a point or a bounding box; a representative coordinate is taken as the point or the box center, and a 1024x1024 candidate region is clipped to image bounds. Predict (P) then feeds the cropped high-resolution patch, optionally with the downsampled global image, to the MLLM for the final answer. The load-bearing assumption is that the MLLM's coarse localization on downsampled images is reliable enough to place the candidate region over the instruction-relevant content, which the random-sampling ablation confirms by collapsing performance when replaced with chance.

What would settle it

Measure the hit rate of the candidate region: run ECP on ScreenSpot-Pro and HR-Bench, record whether the ground-truth target falls inside the 1024x1024 crop, and compare accuracy against a version where the crop is chosen uniformly at random; if the ECP crop's hit rate is not meaningfully above chance for a given base MLLM, the central claim fails for that model.

Watch

Extended reading notes

Core claim

The central claim is that high-resolution MLLM failures can be turned into a coarse-to-fine success: the model's own downsampled-image output still knows roughly where to look. ECP operationalizes this by running the MLLM once on the downsampled full image to extract a candidate point or bounding box, cropping a fixed-size region around that candidate, and running the final prediction on the crop, optionally alongside the downsampled global image for context. On ScreenSpot-Pro, using OS-Atlas-7B in both stages raises accuracy from 19.1% to 40.4%, an absolute gain of 21.3 points; on HR-Bench, Qwen2-VL-7B improves by 5.8 points at 4K and 5.2 points at 8K, with the fine-grained single-instance subset improving by 9.5 and 10.3 points. The random-sampling ablation, which replaces the extracted candidate with a uniformly random crop, drops accuracy below the single-stage baseline, demonstrating that instruction-guided candidate selection, not cropping alone, is what carries the improvement.

Load-bearing premise

The load-bearing premise is that the MLLM's coarse prediction on a downsampled image reliably points near the instruction-relevant content; if that localization misses, the crop misses and accuracy collapses, as the random-sampling ablation shows.

Editorial extensions

If this is right

On ScreenSpot-Pro, ECP with OS-Atlas-7B in both stages raises grounding accuracy from 19.1% to 40.4%, more than doubling the baseline.
On HR-Bench, ECP gives consistent gains at both 4K (+5.8 points) and 8K (+5.2 points), with the largest absolute gains on fine-grained single-instance questions (+9.5 and +10.3 points).
Using a smaller 2B model for candidate extraction still improves a 7B model's final accuracy, suggesting coarse localization is comparatively cheap.
Replacing the candidate with random sampling drops accuracy below the single-stage baseline, so the instruction-guided candidate extraction is essential, not incidental.
Because the same recipe works for GUI grounding with bounding-box output and perception with point output, the framework is task-agnostic by construction.

Reading between the lines

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

If the coarse-to-fine pattern generalizes, any MLLM with even weak localization ability could be wrapped in ECP for domains like document OCR, medical imaging, or satellite imagery, provided the crop size is adapted to the target scale.
The fixed 1024x1024 crop will likely fail when the instruction-relevant content is larger than the crop or spread across multiple distant regions; adaptive crop sizes or multiple candidate crops are natural extensions the paper leaves open.
The paper does not report how often the candidate region actually contains the target; a direct measurement of that hit rate would predict which base models and tasks benefit most from ECP.
ECP could be combined with existing patch-aggregation methods, using ECP to select where to look and aggregation to handle multiple candidate regions within one image.

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.

Desk Editor's Note

read the letter

Straight to it: this paper proposes ECP, a two-stage, training-free scheme where an MLLM first points at a candidate region on a downsampled image, then answers from a 1024x1024 crop. The idea isn't new—iterative coarse-to-fine prompting exists in GUI grounding, and patch-ensemble methods are close—but the task-agnostic framing and the 4K/8K perception results are new, and the paper is written plainly and honestly.

What it does well: the random-sampling ablation is a genuine sanity check; replacing the guided candidate with a uniform random crop collapses accuracy (e.g., OS-Atlas-7B from 40.4% to 11.9% on ScreenSpot-Pro). That tells you the guided selection is doing real work, not just the crop itself. The paper also releases code and reports consistent gains across two benchmarks and multiple model pairings.

Soft spots, in order of seriousness. First, there is no coordinate mapping from the cropped prediction back to the original image, which is required for GUI grounding accuracy to be well-defined. The evaluation protocol must assume some mapping, but it's never described. Second, on HR-Bench there is no comparison to Divide, Conquer and Combine [20], the most directly relevant training-free method; the paper only compares to a downsampled single-stage baseline. Third, the fixed 1024x1024 crop size has no sensitivity analysis, and there are no error bars. Fourth, 'task-agnostic' is slightly oversold, because the perception variant adds the global downsampled image as an extra input, a task-specific design choice.

The stress-test note about the unmeasured hit rate is the right question to ask, but I don't think it sinks the paper. On GUI grounding, stage 2 sees only the crop, with no global image, and accuracy still jumps from 19% to 40% when the model selects the crop. That improvement is indirect evidence that the candidate often contains the target. A direct hit-rate measurement would be a nice addition, but the random-sampling control already rules out the main confounds.

This is a useful incremental paper. It deserves a serious referee. I'd recommend sending it to review, asking for the missing mapping and baselines.

Referee Report

4 major / 4 minor

Summary. The paper introduces Extract Candidate then Predict (ECP), a training-free, task-agnostic two-stage framework for high-resolution image inputs. In Stage 1 (EC), an MLLM is fed a downsampled version of the high-resolution image and asked to output a point or bounding box identifying the instruction-relevant region. In Stage 2 (P), a fixed-size (1024×1024) crop centered on that region is extracted from the original high-resolution image; for perception tasks a global downsampled image is also provided, and the MLLM produces the final answer from this combined input. The framework is evaluated on ScreenSpot-Pro (4K GUI grounding) and HR-Bench (4K/8K MLLM perception), reporting absolute gains of +21.3%, +5.8%, and +5.2% over a single-stage downsampled baseline, together with a random-sampling ablation that shows the guided crop is important.

Significance. If the reported gains are reproducible, ECP is a useful plug-in: it requires no training or per-task heuristics, works with off-the-shelf MLLMs, and the random-sampling ablation is a well-designed control showing that the model-selected crop carries signal. The paper also releases code, which supports reproducibility. The main limitation is that the central mechanism—coarse localization reliability—is asserted rather than directly measured, and several implementation details (coordinate mapping, crop-size sensitivity, comparisons with existing training-free methods) are missing. For these reasons the result is promising but not yet fully established.

major comments (4)

[§3.1 and Table 1] The core premise is that Stage 1 retains coarse localization cues, but the paper never measures whether the candidate region actually contains the instruction-relevant target. The random-sampling ablation only shows that an instruction-guided crop outperforms a uniformly random crop; it does not report the hit rate of Orep inside the ground-truth box, the containment rate of B_candidate, or the distance between Orep and the target. Since the framework's mechanism depends on this premise, please add a direct localization-quality analysis (e.g., hit rate by category and by EC model, and containment rate of the 1024×1024 crop) and compare against a fixed center-crop baseline to rule out center bias.
[§3.1–§3.2, Eqs. (1)–(4), Fig. 2] The GUI grounding evaluation does not specify how Stage-2 outputs are mapped back to the original 4K image. If the final bounding box is produced in the coordinate frame of the cropped patch, it must be shifted by (xleft, ytop) before applying the ScreenSpot-Pro accuracy metric; otherwise the numbers in Table 1 are not comparable to the single-stage baseline. The paper also does not state whether the Stage-1 point or box from the downsampled input is in original-image coordinates or downsampled coordinates, and how any scaling is performed. This is necessary for reproducibility and for the correctness of the grounding results.
[§3.2 and Table 2] In the MLLM perception experiments, Stage 2 always receives the global downsampled image together with the crop, so the improvement over the single-stage baseline could partly come from the model seeing two views rather than from the selected crop. The random-sampling rows show that a random crop plus the global view can be harmful, but they do not isolate the contribution of the model-selected crop. Please add ablations with (i) the global downsampled image presented twice, (ii) a fixed center crop plus the global image, and (iii) an oracle crop guaranteed to contain the target, to confirm that the candidate-selection mechanism drives the observed gains.
[§4, Tables 1–2; §5] The paper compares ECP only with a single-stage downsampled baseline and with random sampling. Prior training-free or iterative methods exist for exactly these benchmarks—DCC [20] for HR-Bench and iterative prompting [9] for ScreenSpot-Pro—so the claim that ECP is general and competitive is not yet supported. Reporting these comparisons (or the numbers already published in [20]) is needed to establish that the task-agnostic design does not lose to existing task-specific pipelines.

minor comments (4)

[§4 and Figures] Table 2's caption contains 'differenct' instead of 'different'; Section 4 contains 'the the single-stage'; and Fig. 1 has a stray 'Incorrect' label near the top of the schematic.
[Eq. (4)] The term 'y bottom' should be typeset as 'y_bottom' for consistency with the other coordinate variables.
[§3.1, §4] The crop size (w,h)=(1024,1024) is the only free hyperparameter but no sensitivity analysis is provided; given that the same size is used for both 4K and 8K images, a short robustness study or at least a motivation for the choice would strengthen the paper.
[Figures 3–4] The qualitative figures would benefit from showing the Stage-1 candidate box overlaid on the full image, so that readers can directly verify the coarse localization step instead of only seeing final results.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: ECP's two-stage pipeline is self-contained and empirically evaluated; the missing hit-rate evidence is a support gap, not a definitional reduction.

full rationale

I walked the paper's claimed derivation chain. The key premise (Sec. 3.1) is that MLLMs retain coarse localization cues on downsampled images. Stage 1 uses the MLLM's own point or bounding-box output as the center of a fixed 1024x1024 crop. Stage 2 runs a separate MLLM inference on the cropped patch, with the global downsampled image also supplied in the perception setting. The reported gains are benchmark accuracies, not quantities forced by construction: no parameter is fitted to the target benchmark outcomes, the crop size (w, h = 1024) is a fixed global hyperparameter, and the candidate region is not defined in terms of the final answer. The random-sampling ablation is a genuine control showing that the identity of the candidate matters, even though it does not directly measure how often the candidate contains the target. That missing hit-rate measurement is an empirical support gap, not circularity. I found no load-bearing self-citation, no imported uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known result. The conclusion's own caveat about limited task coverage is a scope limitation, not evidence that the derivation reduces to its inputs.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The system inherits two assumptions from the base MLLM (coarse localization and coordinate fidelity) plus one hand-chosen hyperparameter (crop size). No new entities are introduced.

free parameters (1)

crop size (w, h) = 1024, 1024
Set to 1024 in all experiments without sensitivity analysis; the candidate bounding box is this size centered on the representative coordinate.

assumptions (3)

domain assumption MLLMs retain coarse localization ability on downsampled high-resolution images
Stated in Section 3.1 and tested via random-sampling ablation; if false, ECP provides no benefit.
domain assumption A single fixed crop size (1024x1024) is sufficient to contain the instruction-relevant content across GUI and perception tasks
If the target is larger than the crop, or the coarse coordinate is off, the cropped patch may exclude the object.
domain assumption The model's stage-1 coordinate output is in the original image coordinate frame and can be directly used to compute the crop
The paper does not describe how stage-2 output coordinates are mapped back to the original image for GUI grounding accuracy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Training-Free, Task-Agnostic Framework for Enhancing MLLM Performance on High-Resolution Images." pith.science (2026). https://pith.science/paper/FR2DA3TE

@misc{pith2026250710202,
  author       = {Pith},
  title        = {Pith review of: A Training-Free, Task-Agnostic Framework for Enhancing MLLM Performance on High-Resolution Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FR2DA3TE}},
  note         = {Machine review of arXiv:2507.10202}
}

read the original abstract

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in vision-language understanding, reasoning, and generation. However, they struggle with tasks requiring fine-grained localization and reasoning in high-resolution images. This constraint stems from the fact that MLLMs are fine-tuned with fixed image resolution to align with the pre-trained image encoder used in MLLM. Consequently, feeding high-resolution images directly into MLLMs leads to poor generalization due to a train-test resolution discrepancy, while downsampling these images-although ensuring consistency-compromises fine-grained visual details and ultimately degrades performance. To address this challenge, we propose Extract Candidate then Predict (ECP), a novel training-free, task-agnostic two-stage framework designed to enhance MLLM performance on high-resolution images. The key intuition behind ECP is that while MLLMs struggle with high-resolution images, their predictions on downsampled images still contain implicit localization cues. By first identifying candidate region using the coarse prediction and then predicting the final output based on candidate region, ECP effectively preserves fine-grained details while mitigating the challenges posed by high-resolution data. We validate our framework on 4K GUI grounding and 4K, 8K MLLM perception, achieving +21.3%, +5.8%, +5.2% absolute improvement compared to baseline respectively, demonstrating its effectiveness. Code is available at https://github.com/yenncye/ECP.

Figures

Figures reproduced from arXiv: 2507.10202 by the authors.

**Figure 1.** (a) Overview of the conventional single-stage framework, where a high-resolution image is first downsampled, and the MLLM generates an output based on downsampled image and text instruction. (b) Overview of our proposed Extract Candidate then Predict (ECP) framework. In the first stage, the MLLM processes a downsampled high-resolution image and generates a candidate region in the form of either a point or a bound… view at source ↗

**Figure 2.** (a) Overview of the ECP framework in GUI Grounding. In the first stage, Extract Candidate (EC), MLLM processes a downsampled high-resolution image and text instruction to generate bounding box for candidate extraction. In the second stage, Predict (P), MLLM generates final prediction based on candidate region and text instruction. (b) Overview of ECP framework in MLLM Perception. In EC, MLLM generates instruction-re… view at source ↗

**Figure 3.** Qualitative results on HR-Bench 8K [20] comparing (a) conventional single-stage framework and (b) the proposed ECP framework. Similarly, on Screenspot-Pro, while the single-stage framework fails to ground the target correctly, our ECP framework identifies its approximate location and provides accurate answers using the cropped image. Ablation For ablation, when applying random sampling in EC, the performance drop… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Qualitative results on ScreenSpot-Pro [9] comparing (a) conventional single-stage framework and (b) the proposed ECP framework. resolution vision-language tasks. We validate the effectiveness of ECP on 4K GUI grounding and 4K, 8K MLLM perception. While our framework i…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

LookWise: Knowing When and Where to Look for Fine-Grained Visual Reasoning in Multimodal Large Language Models
cs.CV 2026-02 conditional novelty 6.0 of 10

A confidence-based when-to-look gate plus semantic-guided attention where-to-look module improves training-free fine-grained visual reasoning accuracy and reduces inference cost versus search-based baselines.

Reference graph

Works this paper leans on

24 extracted references · 9 canonical work pages · cited by 1 Pith paper

[20]

Divide, conquer and combine: A training-free framework for high-resolution im- age perception in multimodal large language models

Wenbin Wang, Liang Ding, Minyan Zeng, Xiabin Zhou, Li Shen, Yong Luo, and Dacheng Tao. Divide, conquer and combine: A training-free framework for high-resolution im- age perception in multimodal large language models. arXiv preprint arXiv:2408.15556, 2024. 1, 2, 4, 5

arXiv 2024
[9]

Screenspot-pro: Gui grounding for professional high- resolution computer use, 2025

Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, and Tat-Seng Chua. Screenspot-pro: Gui grounding for professional high- resolution computer use, 2025. 1, 2, 4, 6

work page 2025
[1]

Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond

Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 2, 4, 5

arXiv 2023
[2]

Lan- guage models are few-shot learners

Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Lan- guage models are few-shot learners. Advances in neural in- formation processing systems, 33:1877–1901, 2020. 2

1901
[3]

Emerg- ing properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 2

2021
[4]

Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks

Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 1, 2

work page 2024
[5]

Seeclick: Har- nessing gui grounding for advanced visual gui agents

Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yan- tao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Har- nessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935, 2024. 4

arXiv 2024
[6]

Palm: Scaling language modeling with pathways

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1–113, 2023. 2

work page 2023

Show all 24 references

[7]

Llava-uhd: an lmm perceiving any aspect ratio and high- resolution images, 2024

Zonghao Guo, Ruyi Xu, Yuan Yao, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high- resolution images, 2024. 2

2024
[8]

Gpt-4o system card

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel- man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli- hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 2, 4, 5

2024 arXiv
[10]

Visual instruction tuning

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892–34916, 2023. 1, 2

2023
[11]

Infimm-hd: A leap forward in high-resolution multimodal understanding

Haogeng Liu, Quanzeng You, Xiaotian Han, Yiqi Wang, Bo- han Zhai, Yongfei Liu, Yunzhe Tao, Huaibo Huang, Ran He, and Hongxia Yang. Infimm-hd: A leap forward in high-resolution multimodal understanding. arXiv preprint arXiv:2403.01487, 2024. 2

2024 arXiv
[12]

Feast your eyes: Mixture-of- resolution adaptation for multimodal large language models

Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xi- aoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of- resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024. 2

2024 arXiv
[13]

Inf-llava: Dual- perspective perception for high-resolution multimodal large language model

Yiwei Ma, Zhibin Wang, Xiaoshuai Sun, Weihuang Lin, Qiang Zhou, Jiayi Ji, and Rongrong Ji. Inf-llava: Dual- perspective perception for high-resolution multimodal large language model. arXiv preprint arXiv:2407.16198, 2024. 2

2024 arXiv
[14]

Learning transferable visual models from natural language supervi- sion

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

2021
[15]

Divide and conquer: High-resolution industrial anomaly de- tection via memory efficient tiled ensemble

Bla ˇz Rolih, Dick Ameln, Ashwin Vaidya, and Samet Akcay. Divide and conquer: High-resolution industrial anomaly de- tection via memory efficient tiled ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 3866–3875, 2024. 2

2024
[16]

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of con- text

Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of con- text. arXiv preprint arXiv:2403.05530, 2024. 1, 2

2024 arXiv
[17]

Fixing the train-test resolution discrepancy.Advances in neural information processing systems, 32, 2019

Hugo Touvron, Andrea Vedaldi, Matthijs Douze, and Herv ´e J´egou. Fixing the train-test resolution discrepancy.Advances in neural information processing systems, 32, 2019. 1

2019
[18]

Llama: Open and efficient foundation language models

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2

2023 arXiv
[19]

Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2

2024 arXiv
[21]

Os-atlas: A foundation action model for generalist gui agents

Zhiyong Wu, Zhenyu Wu, Fangzhi Xu, Yian Wang, Qiushi Sun, Chengyou Jia, Kanzhi Cheng, Zichen Ding, Liheng Chen, Paul Pu Liang, et al. Os-atlas: A foundation action model for generalist gui agents. arXiv preprint arXiv:2410.23218, 2024. 4

arXiv
[22]

Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding

Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. Advances in Neural Informa- tion Processing Systems, 37:71737–71767, 2025. 1

2025
[23]

Large language models are not robust multi- ple choice selectors

Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. Large language models are not robust multi- ple choice selectors. arXiv preprint arXiv:2309.03882, 2023. 4

2023 arXiv
[24]

Mrovseg: Breaking the resolution curse of vision-language models in open-vocabulary seman- tic segmentation

Yuanbing Zhu, Bingke Zhu, Zhen Chen, Huan Xu, Ming Tang, and Jinqiao Wang. Mrovseg: Breaking the resolution curse of vision-language models in open-vocabulary seman- tic segmentation. arXiv preprint arXiv:2408.14776 , 2024. 1 7

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [20]

Divide, conquer and combine: A training-free framework for high-resolution im- age perception in multimodal large language models

Wenbin Wang, Liang Ding, Minyan Zeng, Xiabin Zhou, Li Shen, Yong Luo, and Dacheng Tao. Divide, conquer and combine: A training-free framework for high-resolution im- age perception in multimodal large language models. arXiv preprint arXiv:2408.15556, 2024. 1, 2, 4, 5

arXiv 2024

[2] [9]

Screenspot-pro: Gui grounding for professional high- resolution computer use, 2025

Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, and Tat-Seng Chua. Screenspot-pro: Gui grounding for professional high- resolution computer use, 2025. 1, 2, 4, 6

work page 2025

[3] [1]

Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond

Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 2, 4, 5

arXiv 2023

[4] [2]

Lan- guage models are few-shot learners

Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Lan- guage models are few-shot learners. Advances in neural in- formation processing systems, 33:1877–1901, 2020. 2

1901

[5] [3]

Emerg- ing properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 2

2021

[6] [4]

Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks

Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 1, 2

work page 2024

[7] [5]

Seeclick: Har- nessing gui grounding for advanced visual gui agents

Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yan- tao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Har- nessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935, 2024. 4

arXiv 2024

[8] [6]

Palm: Scaling language modeling with pathways

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1–113, 2023. 2

work page 2023