Pith. sign in

REVIEW 4 major objections 6 minor 80 references

A 0.9B screen-reading model can outperform 2B general VLMs on UI tasks and still run on a phone after 4-bit quantization.

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 16:27 UTC pith:VRP2V7XB

load-bearing objection A credible on-device UI VLM engineering recipe whose headline deployment claim—quantization loss within 1%—is asserted but not evidenced; the missing quantized-vs-FP accuracy table is the main gap. the 4 major comments →

arxiv 2607.22708 v1 pith:VRP2V7XB submitted 2026-07-20 cs.CV

StepX-Edge: An On-Device UI Vision-Language Model via Architecture-Training-Deployment Co-Design

Yin Wang , Haotian Hu , Jineng Han , Wentao Qiu , Zhenhua Ge , Liujian Tang , Fanyi Wang This is my paper
classification cs.CV
keywords on-device vision-language modelUI understandingOCRelement groundingscreen VQAcurriculum learningquantizationmobile deployment
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 tries to prove that a small, phone-runnable vision-language model can beat much larger general-purpose models at understanding user interfaces, provided architecture, training data, and quantization are co-designed rather than bolted together. It reports a 0.9B model that tops all baselines, including 2B-2.3B competitors, on screen question answering (F1 88.76) and Chinese OCR (57.25), and matches larger models on element grounding and general OCR. The explanation rests on a claimed mutual-reinforcement effect among OCR, grounding, VQA, and screen understanding, exploited through a five-stage curriculum that teaches grounding first. A deployment scheme combining 8-bit vision with 4-bit language quantization keeps accuracy loss under 1% and delivers 0.84 s first token, 98 tokens/s, and 1.4 GB memory on a phone chip. If these results hold, vertical specialization plus careful co-design can substitute for raw scale in on-device UI agents.

Core claim

The paper sets out to show that a 0.9B-parameter vision-language model built specifically for mobile UI screens can outperform 2B-2.3B general-purpose baselines on screen-focused tasks—ScreenQA F1 88.76 versus about 75 for the next best model, and Chinese OCRBench v2 57.25 versus 53.0 for the best 2B model—while matching larger models on RefCOCO grounding (92.0%) and general OCR (831). It further claims that after module-wise quantization (W8A16 for vision, W4A16+KV8 for the language model), the model runs end-to-end on a Snapdragon 8 Gen5 device at 0.84 s time-to-first-token, 98 tokens/s, and 1.4 GB peak memory, with post-quantization accuracy loss within 1%. The evidence is organized aroun

What carries the argument

The load-bearing mechanism is the claimed mutual-reinforcement loop among UI subtasks: grounding forces precise spatial encoding, which sharpens OCR's character and line localization; OCR's accumulated textual semantics then anchor grounding, all within a shared, frozen-vision-encoder representation during the first two curriculum stages. Supporting it are ULVE (adaptive tiling plus channel-attention feature fusion for extreme aspect ratios), PDP (a three-layer MLP that refines the 2048-dim visual space, compresses to 1024, then adapts to the language space), and a two-stage PTQ-to-QAT quantization pipeline that freezes the vision side while the language model is trained to compensate for 4-

Load-bearing premise

The load-bearing premise is that the comparison numbers are fair: every baseline was run from official weights with prompts and resolutions aligned, and no benchmark test image leaked into training via the pHash deduplication; if any of those conditions fails, the reported leads—especially the 13-point ScreenQA gap—are not established.

What would settle it

Run an independent, third-party evaluation of StepX-Edge and the same baselines on ScreenQA and OCRBench v2 using a fixed harness with pre-registered prompts and resolutions, and verify with pHash that no test image is in the training pool; if the ScreenQA lead of 13 points or the Chinese OCRBench v2 lead of about 4 points disappears, the central superiority claim fails.

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

If this is right

  • On-device UI agents become practical: a real phone can run the model at about 98 tokens per second with 1.4 GB peak memory and about 0.84 s to the first token.
  • A specialized 0.9B model can beat 2B-2.3B general VLMs on screen QA and Chinese OCR, suggesting vertical specialization can compensate for scale in narrow domains.
  • The five-stage curriculum with grounding trained first is proposed as a reusable recipe for small multimodal models that must handle several fine-grained tasks at once.
  • The quantitative evidence shows the strongest gains are in screen reading and Chinese OCR; on mobile interaction grounding (ScreenSpot v2) a 2B baseline still leads 91.0 vs 72.5, so 'strongest overall' is not universal.
  • The open-sourced data, training recipe, and deployment pipeline imply the result can be independently reproduced and adapted rather than remaining a one-off system.

Where Pith is reading between the lines

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

  • The paper does not ablate the mutual-reinforcement claim directly; a clean test would train the same 0.9B architecture on each task alone and on the full mixture, then check whether joint training beats single-task training on ScreenQA and OCR.
  • The 'refine before reduction, adapt after reduction' connector principle is likely transferable to other on-device multimodal settings where a large visual embedding dimension meets a smaller language dimension, such as document understanding or sign recognition.
  • The fixed 768x1344 input and full-attention choice imply a deliberate resolution-throughput trade-off; reducing visual tokens would raise serving throughput but likely cut grounding accuracy, a knob the paper documents but leaves for future hardware.
  • The observation of massive outlier activations in the vision encoder suggests that activation-aware quantization, not just weight quantization, is the main risk for UI-focused VLMs; future work could explicitly target those outlier channels in the pre-quantization calibration.

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 / 6 minor

Summary. StepX-Edge is a 0.9B-parameter on-device UI vision-language model combining a Qwen3.5-ViT encoder adapted with UI-Aware Layered Visual Encoding (ULVE), a three-layer Progressive Dimensionality Projection (PDP) connector, and a Qwen3-0.6B language decoder. The training pipeline is a five-stage curriculum (modality alignment, multi-task pretraining, foundational SFT, advanced SFT, RL alignment) motivated by an asserted mutual-reinforcement effect among OCR, screen understanding, VQA, and UI grounding. Deployment uses module-wise differentiation: W8A16 for ViT/connector and W4A16+KV8 for the LLM, with a two-stage PTQ-to-QAT procedure. The paper reports state-of-the-art scores for a ≤1B model on ScreenQA (88.76 F1) and Chinese OCRBench v2 (57.25), competitive results on RefCOCO (92.0) and OCRBench v1 (831), and on-device runtime of ~0.84 s TTFT, ~98 tok/s decode, and ~1.4 GB peak memory on Snapdragon 8 Gen5, while claiming post-quantization accuracy loss within 1%.

Significance. If substantiated, this would be a useful systems contribution: a real-device-deployed, low-bit quantized small UI VLM with a reproducible training recipe, real hardware measurements rather than simulation, and an explicit pHash-based test-set deduplication step. The paper also gives concrete GPU server-side runtime numbers and describes a hardware-aware architecture (full attention, standard operators). The main value is the complete accuracy-runtime picture. However, the central quantitative claims are currently unevenly supported: the deployment claim lacks the FP-vs-quantized accuracy measurement, the curriculum claim lacks any ablation, and the baseline-comparison protocol is described only at a high level. These are fixable with additional experiments and documentation, but they are load-bearing for the abstract's promises.

major comments (4)
  1. [§2.4, §3.4, Table 7] The abstract, §1, and §2.4.2 state that W4A16+KV8 quantization keeps post-quantization accuracy loss within 1%, and §2.4.2 concludes that 'the target of an accuracy loss within 1% is achieved.' No measurement is shown. Table 7 and §3.4 report only runtime/memory numbers; there is no quantized-vs-full-precision comparison on ScreenQA, OCRBench v1/v2, RefCOCO, or ScreenSpot. The claim is therefore unverifiable as written. Please add a per-benchmark FP-vs-quantized table and define whether '1%' is average or worst-case, absolute or relative. This is directly load-bearing for the abstract's deployment claim, especially since §2.4 argues UI tasks are sensitive to low-bit quantization.
  2. [§2.3, §3.2] The paper's second contribution is the 'discovery and validation' of mutual reinforcement among OCR, screen understanding, VQA, and grounding, and the five-stage curriculum is built on this. No experiment supports this claim: there is no ablation that trains tasks separately, removes the curriculum, or varies the 'engine-task-first' weighting. The reader only sees the final system's results. The mutual-reinforcement assertion is central to the training design and should be tested by at least one ablation (e.g., flat multi-task mixing vs. the proposed curriculum, or joint vs. isolated task training). Without it, the contribution is an interpretation rather than a validated finding.
  3. [§3.1] All head-to-head comparisons rest on the statement that baselines were run 'using the official open-source weights and default inference configuration, with input resolution and prompt templates fully aligned with StepX-Edge.' The manuscript does not report the exact resolutions, prompt templates, decoding settings, or number of runs for each baseline, and no evaluation code is provided. These choices can move scores by several points on OCR and grounding benchmarks, so the claimed superiority on ScreenQA and OCRBench v2 is not independently checkable from the paper. Please include a per-baseline configuration table and error bars or run-to-run variance, and release the evaluation harness.
  4. [§3.2, ScreenSpot v2 Mobile] The phrase 'strongest overall UI-understanding performance' is used without a defined aggregation. StepX-Edge trails Qwen3-VL-2B on ScreenSpot v2 Mobile (72.51 vs 91.02) and is essentially tied on OCRBench v2 en (55.03 vs 56.2), while leading on ScreenQA and Chinese OCRBench v2. If 'overall' refers to a specific normalized average, the formula should be stated; otherwise the per-benchmark claims should be qualified. This will prevent over-reading of the abstract's 'strongest overall' wording.
minor comments (6)
  1. [§2.2.2] The curriculum mixing weights (e.g., 0.25 for cn_mobile_grounding, 0.2 for seeclick_hard) are described as 'determined through multiple rounds of iteration,' but no sensitivity analysis is reported. A small range study would strengthen the reproducibility of the recipe.
  2. [§2.4.2] The QAT calibration set is described as 'about 512 batches'; it is unclear whether this is 512 samples or 512 batches, and no QAT iteration count or loss is given. Please clarify.
  3. [§3.1] OCRBench v2 is said to take the 'average of the official metrics for each subset,' but the official metrics are not enumerated. Please specify which metrics are averaged.
  4. [Figure 3] Each panel uses the label 'Overall (higher is better),' which is confusing because it is not an overall score but the score of the specified benchmark. Rename the y-axis labels accordingly.
  5. [§1, §2.1.3] There are typographical artifacts such as 'T wo' and the repeated use of 'efficiency' (non-ASCII ligature) in the rendered text. Please proofread.
  6. [§2.1.1, §2.1.2] The ULVE and PDP components are not ablated. An ablation table (e.g., removing adaptive tiling, replacing PDP with a single-layer MLP or cross-attention) would make the architecture claims more than post-hoc attributions in §3.2.1.

Circularity Check

0 steps flagged

No reduction-by-construction found; central benchmark claims rest on external benchmarks. Minor post-hoc 'mutual reinforcement' narrative and an unmeasured 1% quantization-loss claim are evidence gaps, not formal circularity.

full rationale

StepX-Edge's headline results (ScreenQA 88.76, OCRBench v2 cn 57.25, RefCOCO 92.0, OCRBench v1 831) are measured against external public benchmarks with independent baselines; no fitted parameter is renamed as a prediction, and no equation in the paper equates an output to an input. Eq. (1) (channel-attention fusion) and Eq. (2) (reward combination) are design definitions, not derivations that reduce to their own inputs. The paper contains no load-bearing self-citation or imported uniqueness theorem; references to prior work are standard external sources. Two weaknesses are noted but are not circularity under the stated definition. First, §2.4.2 states 'the target of an accuracy loss within 1% is achieved' while §3.4 reports only runtime metrics (TTFT, decode speed, peak memory) with no quantized-versus-floating-point accuracy table; this is a missing-evidence/correctness risk, not a circular derivation. Second, the 'mutual reinforcement' effect (§2.3) is inferred from the authors' own training runs and then treated as validated by the final model's success in §3.2.1 ('consistent with multi-task mutually-reinforcing training'); this is post-hoc and lacks a controlled ablation, making it self-referential in an evidentiary sense, but it does not construct any numerical prediction from its inputs. Accordingly, no significant formal circularity is present; the score reflects the minor internal-consistency/evidence gaps rather than any reduction-by-construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper is an empirical systems contribution: headline claims rest on self-reported fair evaluation and leakage prevention, and the design claims (mutual reinforcement, curriculum benefit, <1% quantization loss) are asserted without ablations or a quantization accuracy table. Many hyperparameters are hand-selected. No new physical entities are introduced.

free parameters (6)
  • Curriculum general-data retention weights = 1.0 → 0.2–0.3 → 10–15%
    Set by iteration in §2.2.2, stated as 'determined through multiple rounds of iteration'; directly shapes the capability balance.
  • Stage peak learning rates and ViT LR factor = 1e-4, 2e-5, 1e-5, 5e-6; ViT 1/10 of LLM LR
    Chosen per stage in Table 4; no sensitivity analysis or sweeps reported.
  • Mid-training engine-task weights = cn_mobile_grounding 0.25, seeclick_hard 0.2, overall range 0.1–1.0
    Grounding-first mixing in §2.2.2 is hand-selected; no evidence shown that these weights are optimal.
  • RL reward weights = (0.5, 0.2, 0.1, 0.2)
    Eq. (2) in §2.3.5; chosen values with no reported ablation.
  • Quantization bit-width choices = ViT W8A16, LLM W4A16+KV8, layer_2.down_proj 8-bit, embedding 8-bit
    Module-wise choices in §2.4.1 motivated by measured outliers and bandwidth; no comparison of alternative bit-widths.
  • Calibration and QAT hyperparameters = 1024 calibration images, 512 batches, QAT LR 2e-6
    Hand-picked in §2.4.2; only one configuration is reported.
axioms (5)
  • domain assumption Joint training of OCR, screen understanding, VQA, and grounding in a shared frozen-ViT feature space produces mutual reinforcement rather than interference.
    Invoked in §2.3.2 as the basis for the curriculum; no controlled ablation against naive multi-task mixing is reported.
  • domain assumption Freezing ViT main layers preserves pretrained visual knowledge while making the Merger trainable early accelerates UI-domain adaptation.
    Assumed in §2.1.1 and §2.3.6; no layer-wise ablation or feature-distance evidence.
  • domain assumption The self-reported data-cleaning and pHash leakage checks remove all evaluation-benchmark overlap from training data.
    Load-bearing for benchmark validity; all headline numbers depend on §2.2.3 and cannot be independently checked because the data is not released.
  • domain assumption QAT on the LLM alone, with a frozen W8A16 ViT front-end, restores 4-bit quantization loss to within 1%.
    Stated in §2.4.2; no quantized-vs-bf16 accuracy table is provided.
  • domain assumption Standard full attention is efficiently supported by current mobile NPU/HTP operator stacks.
    Used in §2.1.3 and §2.4.1 to justify the architecture; plausible but not benchmarked against alternative attention variants.

pith-pipeline@v1.3.0-alltime-deepseek · 28248 in / 13004 out tokens · 127231 ms · 2026-08-01T16:27:10.536960+00:00 · methodology

0 comments
read the original abstract

Deploying a vision-language model with full UI understanding on end devices has long been trapped between accuracy and efficiency: on one side is the accuracy bar for OCR, screen understanding, visual question answering, and element grounding; on the other is the strict compute, memory, and power budget of mobile chips. Existing work either trades one for the other, or stops at simulation without real-device validation. We present StepX-Edge, a 0.9B-parameter on-device UI vision-language model that resolves this tension through three-layer co-design of architecture, training, and deployment. Architecturally, UI-aware Layered Visual Encoding (ULVE) and a Progressive Dimensionality Projection (PDP) connector target the extreme aspect ratios and fine-grained perception of screens, while standard full attention throughout ensures native compatibility with mainstream mobile NPU operators. For training, the five-stage StepX-Curriculum framework is designed around our observation of mutual-promotion effects among UI subtasks, so that all four capabilities grow synergistically under a tight parameter budget rather than interfering. For deployment, a module-wise differentiated two-stage PTQ-to-QAT quantization scheme keeps the post-quantization accuracy loss within 1%. StepX-Edge achieves the strongest overall UI understanding among <=1B models, surpassing all 2B-2.3B baselines on ScreenQA (88.76 F1) and Chinese OCRBench v2 (57.25), and matching 1.3B-2.3B general VLMs on RefCOCO (92.0%) and OCRBench v1 (831) with far fewer parameters. After W4A16+KV8 quantization, the model runs stably on Snapdragon 8 Gen5 devices with ~0.84 s TTFT, 98 tok/s decode, and 1.4 GB peak memory. We will open-source the training data, the full training recipe, and the quantization deployment pipeline.

Figures

Figures reproduced from arXiv: 2607.22708 by Fanyi Wang, Haotian Hu, Jineng Han, Liujian Tang, Wentao Qiu, Yin Wang, Zhenhua Ge.

Figure 1
Figure 1. Figure 1: Overall architecture of StepX-Edge and its on-device deployment pipeline. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The five-stage progressive training pipeline of StepX-Curriculum. Top: the trainable modules, [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of StepX-Edge with mainstream on-device VLMs across UI-understanding bench [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: GPU server-side runtime comparison of StepX-Edge with mainstream on-device VLMs [PITH_FULL_IMAGE:figures/full_fig_p025_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: OCR case (English handwriting scenario): a handwritten manuscript of the lyrics of Yesterday Once More. The model transcribes the title, the Chinese translated name, and all lyrics line by line, and, drawing on semantic context, restores the handwritten strokes into a well-formed song title and lyrics. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: UI screen-understanding case (Chinese scenario): a post-detail page of a note-sharing app (a classical-music recommendation post). The model correctly outputs the semantic sections and the interaction counts. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: UI screen-understanding case (English scenario): the Categories page of the Sport Live TV app. The model recognizes the tab bar, the banner, and the category entries. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: VQA case (Chinese scenario): a product quality test report. The question requires multi￾step reasoning—computing the day gap between two dates, counting the reference-document entries, and reading the pass/fail conclusion—which the model answers correctly. 42 [PITH_FULL_IMAGE:figures/full_fig_p042_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: VQA case (English scenario): a recipe detail page. The model aggregates three categories of structured information—duration, servings, and ingredients. 43 [PITH_FULL_IMAGE:figures/full_fig_p043_10.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

80 extracted references · 23 linked inside Pith

  1. [1]

    Flamingo: A Visual Language Model for Few-Shot Learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, and Y ana Hasson et al. Flamingo: A Visual Language Model for Few-Shot Learning. Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  2. [2]

    Visual Instruction Tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Y ong Jae Lee. Visual Instruction Tuning. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  3. [3]

    GPT-4V(ision) System Card

    OpenAI. GPT-4V(ision) System Card. https://openai.com/index/gpt-4v-system-card/ , 2023

    2023

  4. [4]

    Qwen2.5-VL Technical Report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, and Humen Zhong et al. Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipes

    Tianyu Yu, Zefan Wang, Chongyi Wang, Fuwei Huang, Wenshuo Ma, Zhihui He, Tianchi Cai, Weize Chen, and Yuxiang Huang et al. MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipes. arXiv preprint arXiv:2509.18154, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Gemma Team, Sherif El Abd, Vaibhav Aggarwal, Robin Algayres, Alek Andreev, Olivier Bachem, Ian Ballantyne, Cormac Brick, Victor Cărbune, Michelle Casbon, Mayank Chaturvedi, Victor Cotruta, Alice Coucke, Phil Culliton, Robert Dadashi, Lucas Dixon, Mohamed Elhawaty, Utku Evci, Clément Farabet, Johan Ferret, Filippo Galgani, Sertan Girgin, Jean-Bastien Grill...

    Pith/arXiv arXiv 2026

  7. [7]

    Qwen3.5-Omni Technical Report

    Qwen Team. Qwen3.5-Omni Technical Report. arXiv preprint arXiv:2604.15804 , 2026. Vision capability of Qwen3.5-VL-0.8B is provided through the natively multimodal Qwen3.5-Omni model

    Pith/arXiv arXiv 2026

  8. [8]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, and Lucile Saulnier et al. Mistral 7B. arXiv preprint arXiv:2310.06825, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Generating Long Sequences with Sparse Transformers

    Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating Long Sequences with Sparse Transformers. arXiv preprint arXiv:1904.10509, 2019

    Pith/arXiv arXiv 1904

  10. [10]

    GPTQ: Accurate Post-Training 29 Quantization for Generative Pre-Trained Transformers

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. GPTQ: Accurate Post-Training 29 Quantization for Generative Pre-Trained Transformers. In International Conference on Learning Representations (ICLR), 2023

    2023

  11. [11]

    A WQ: Activation-Aware Weight Quantization for LLM Compression and Acceleration

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Y ang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. A WQ: Activation-Aware Weight Quantization for LLM Compression and Acceleration. In Conference on Machine Learning and Systems (MLSys) , 2024

    2024

  12. [12]

    SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models. In International Conference on Machine Learning (ICML) , 2023

    2023

  13. [13]

    Qwen3 Technical Report

    Qwen Team. Qwen3 Technical Report. arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Qwen3-VL Technical Report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Y ang Liu, Dayiheng Liu, Shixuan...

    Pith/arXiv arXiv 2025

  15. [15]

    LLaV A-UHD: An LMM Perceiving Any Aspect Ratio and High-Resolution Images

    Ruyi Xu, Yuan Y ao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. LLaV A-UHD: An LMM Perceiving Any Aspect Ratio and High-Resolution Images. In European Conference on Computer Vision (ECCV) , 2024

    2024

  16. [16]

    Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang

    Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P . Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  17. [17]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is All Y ou Need. In Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  18. [18]

    ICDAR 2019 Robust Reading Challenge on Large-Scale Street View Text with Partial Labeling (LSVT)

    Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, and Dimosthenis Karatzas et al. ICDAR 2019 Robust Reading Challenge on Large-Scale Street View Text with Partial Labeling (LSVT). In International Conference on Document Analysis and Recognition (ICDAR) , 2019. 30

    2019

  19. [19]

    MTWI: A Large Multi-T ype Web Images Dataset for Text Recognition

    Meng He, Yuliang Liu, Zhenbo Y ang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Y ongpan Wang, Xin Zhang, and Lianwen Jin. MTWI: A Large Multi-T ype Web Images Dataset for Text Recognition. arXiv preprint arXiv:1810.03526, 2018

    Pith/arXiv arXiv 2018

  20. [20]

    COCO-Text: Dataset and Benchmark for Text Detection and Recognition in Natural Images

    Andreas Veit, Tomas Matera, Lukáš Neumann, Jiří Matas, and Serge Belongie. COCO-Text: Dataset and Benchmark for Text Detection and Recognition in Natural Images. arXiv preprint arXiv:1601.07140, 2016

    Pith/arXiv arXiv 2016

  21. [21]

    Synthetic Data for Text Localisation in Natural Images

    Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic Data for Text Localisation in Natural Images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2016

    2016

  22. [22]

    Tex- tOCR: Towards Large-Scale End-to-End Reasoning for Arbitrary-Shaped Scene Text

    Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. Tex- tOCR: Towards Large-Scale End-to-End Reasoning for Arbitrary-Shaped Scene Text. InIEEE Con- ference on Computer Vision and Pattern Recognition (CVPR) , 2021

    2021

  23. [23]

    Jawa- har

    Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C.V . Jawa- har. ICDAR 2019 Competition on Scanned Receipt OCR and Information Extraction. In Interna- tional Conference on Document Analysis and Recognition (ICDAR) , 2019

    2019

  24. [24]

    XFUND: A Benchmark Dataset for Multilingual Visually Rich Form Understanding

    Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, and Furu Wei. XFUND: A Benchmark Dataset for Multilingual Visually Rich Form Understanding. Findings of ACL, 2022

    2022

  25. [25]

    FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents

    Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents. In ICDAR Workshops, 2019

    2019

  26. [26]

    Image-Based Table Recognition: Data, Model, and Evaluation (PubTabNet)

    Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Y epes. Image-Based Table Recognition: Data, Model, and Evaluation (PubTabNet). In European Conference on Computer Vision (ECCV) , 2020

    2020

  27. [27]

    Spatial Dual- Modality Graph Reasoning for Key Information Extraction (WildReceipt)

    Hongbin Sun, Zhanghui Kuang, Xiaoyu Yue, Chenhao Lin, and Wayne Zhang. Spatial Dual- Modality Graph Reasoning for Key Information Extraction (WildReceipt). arXiv preprint arXiv:2103.14470, 2021

    Pith/arXiv arXiv 2021

  28. [28]

    CASIA Online and Offline Chinese Handwriting Databases

    Cheng-Lin Liu, Fei Yin, Da-Han Wang, and Qiu-Feng Wang. CASIA Online and Offline Chinese Handwriting Databases. International Conference on Document Analysis and Recognition (ICDAR), 2011

    2011

  29. [29]

    Synthetic Data and Ar- tificial Neural Networks for Natural Scene Text Recognition.NeurIPS Workshop on Deep Learning, 2014

    Max Jaderberg, Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Synthetic Data and Ar- tificial Neural Networks for Natural Scene Text Recognition.NeurIPS Workshop on Deep Learning, 2014

    2014

  30. [30]

    Industry Documents Library WebDataset (IDL-WDS)

    Industry Documents Library. Industry Documents Library WebDataset (IDL-WDS). https:// huggingface.co/datasets/pixparse/idl-wds, 2023. 31

    2023

  31. [31]

    Towards End-to-End Unified Scene Text Detection and Layout Analysis (HierText)

    Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Y asuhisa Fujii, and Michalis Raptis. Towards End-to-End Unified Scene Text Detection and Layout Analysis (HierText). IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2023

    2023

  32. [32]

    SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents

    Kanzhi Cheng, Qiushi Sun, Y ougang Chu, Fangzhi Xu, Y antao Li, Jianbing Zhang, and Zhiyong Wu. SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents. In Annual Meeting of the Association for Computational Linguistics (ACL) , 2024

    2024

  33. [33]

    ShowUI: One Vision-Language-Action Model for GUI Visual Agent

    Kevin Qinghong Lin, Linjie Li, and Difei Gao et al. ShowUI: One Vision-Language-Action Model for GUI Visual Agent. arXiv preprint arXiv:2411.17465, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Wave-UI-25K: A Large-Scale UI Element Dataset

    Agentsea. Wave-UI-25K: A Large-Scale UI Element Dataset. https://huggingface.co/ datasets/agentsea/wave-ui-25k, 2024

    2024

  35. [35]

    MobileViews: A Large-Scale Mobile GUI Dataset

    MobileViews Team. MobileViews: A Large-Scale Mobile GUI Dataset. https://github.com/ mobile-viz/MobileViews, 2024

    2024

  36. [36]

    Screen2Words: Automatic Mobile UI Summarization with Multimodal Learning

    Bryan Wang, Gang Li, Xin Zhou, Zhourong Chen, Tovi Grossman, and Y ang Li. Screen2Words: Automatic Mobile UI Summarization with Multimodal Learning. In ACM Symposium on User Interface Software and Technology (UIST), 2021

    2021

  37. [37]

    On the Effects of Data Scale on UI Control Agents (Android Control)

    Wei Li, William Bishop, Alice Li, Chris Rawles, Folawiyo Campos, Divya Chandra, and Kate Be- rard et al. On the Effects of Data Scale on UI Control Agents (Android Control). Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  38. [38]

    Android in the Wild: A Large-Scale Dataset for Android Device Control

    Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Android in the Wild: A Large-Scale Dataset for Android Device Control. Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  39. [39]

    AMEX: Android Multi-annotation Expo Dataset for Mobile GUI Agents

    Yuxiang Chai, Siyuan Huang, Y azhe Niu, Han Xiao, Liang Liu, Dingyu Yuan, Shuai Gao, Jizhong Zhang, and Hongsheng Sun. AMEX: Android Multi-annotation Expo Dataset for Mobile GUI Agents. arXiv preprint arXiv:2407.17490, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    GUIAct: A Large-Scale Dataset for GUI Action Understand- ing

    Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, and Yupeng Huo et al. GUIAct: A Large-Scale Dataset for GUI Action Understand- ing. arXiv preprint arXiv:2406.10819, 2024

    Pith/arXiv arXiv 2024

  41. [41]

    AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents

    Christopher Rawles, Sarah Clinckemaillie, Yifan Chang, Jonathan Waltz, Gabrielle Lau, Marybeth Fair, Alice Li, William Bishop, Wei Li, and Folawiyo Campos et al. AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents. arXiv preprint arXiv:2405.14573, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    ScreenQA: Large-Scale Question-Answer Pairs Over Mobile App Screen- shots

    Yu-Chung Hsiao, Fedir Zubach, Gilles Baechler, Srinivas Sunkara, Victor Rusu, Jindong Chen, and Abhanshu Rajkumar. ScreenQA: Large-Scale Question-Answer Pairs Over Mobile App Screen- shots. arXiv preprint arXiv:2209.08199, 2022. 32

    Pith/arXiv arXiv 2022

  43. [43]

    Minesh Mathew, Dimosthenis Karatzas, and C.V . Jawahar. DocVQA: A Dataset for VQA on Doc- ument Images. In IEEE Winter Conference on Applications of Computer Vision (WACV) , 2021

    2021

  44. [44]

    Minesh Mathew, Viraj Bagal, Rubèn Tito, Dimosthenis Karatzas, Ernest Valveny, and C.V . Jawahar. InfographicVQA. In IEEE Winter Conference on Applications of Computer Vision (WACV) , 2022

    2022

  45. [45]

    ChineseDocVQA: A Chinese Document Question Answering Benchmark

    Chinese Document VQA Team. ChineseDocVQA: A Chinese Document Question Answering Benchmark. https://huggingface.co/datasets/CADT/ChineseDocVQA, 2023

    2023

  46. [46]

    ChartQA: A Bench- mark for Question Answering about Charts with Visual and Logical Reasoning

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A Bench- mark for Question Answering about Charts with Visual and Logical Reasoning. In Findings of ACL, 2022

    2022

  47. [47]

    Towards VQA Models That Can Read

    Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA Models That Can Read. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  48. [48]

    Jawahar, and Dimosthenis Karatzas

    Ali Furkan Biten, Rubèn Tito, Andres Mafla, Lluis Gomez, Marçal Rusiñol, Ernest Valveny, C.V . Jawahar, and Dimosthenis Karatzas. ST-VQA: Scene Text Visual Question Answering. In Interna- tional Conference on Computer Vision (ICCV) , 2019

    2019

  49. [49]

    OCR-VQA: Visual Question Answering by Reading Text in Images

    Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. OCR-VQA: Visual Question Answering by Reading Text in Images. International Conference on Document Analysis and Recognition (ICDAR), 2019

    2019

  50. [50]

    Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering. In Advances in Neural Information Processing Systems (NeurIPS) , 2022

    2022

  51. [51]

    MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts. In International Conference on Learning Representations (ICLR), 2024

    2024

  52. [52]

    LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image-Text Models.Advances in Neural Information Processing Systems (NeurIPS), 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, and Mitchell Wortsman et al. LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image-Text Models.Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  53. [53]

    Conceptual 12M: Pushing Web- Scale Image-Text Pre-Training to Recognize Long-Tail Visual Concepts

    Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing Web- Scale Image-Text Pre-Training to Recognize Long-Tail Visual Concepts. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2021. 33

    2021

  54. [54]

    Wukong: A 100 Million Large-Scale Chinese Cross-Modal Pre-Training Benchmark

    Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Xiaodan Liang, Lewei Y ao, Runhui Huang, Wei Zhang, and Xin Jiang et al. Wukong: A 100 Million Large-Scale Chinese Cross-Modal Pre-Training Benchmark. Advances in Neural Information Processing Systems (NeurIPS) , 2022

    2022

  55. [55]

    COCO-CN for Cross-Lingual Image Tagging, Captioning, and Retrieval

    Xirong Li, Chaoxi Xu, Xiaoxu Wang, Weiyu Lan, Zhengxiong Jia, Gang Y ang, and Jieping Xu. COCO-CN for Cross-Lingual Image Tagging, Captioning, and Retrieval. IEEE Transactions on Multimedia, 2019

    2019

  56. [56]

    Flickr30k-CN: Chinese Version of Flickr30k

    Weiyu Lan, Xirong Li, and Jianfeng Dong. Flickr30k-CN: Chinese Version of Flickr30k. arXiv preprint arXiv:1709.01754, 2017

    arXiv 2017

  57. [57]

    ALLaV A: Harnessing GPT4V-Synthesized Data for a Lite Vision-Language Model

    Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhi- hong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. ALLaV A: Harnessing GPT4V-Synthesized Data for a Lite Vision-Language Model. arXiv preprint arXiv:2402.11684, 2024

    Pith/arXiv arXiv 2024

  58. [58]

    ShareGPT4V: Improving Large Multi-Modal Models with Better Captions

    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. In European Con- ference on Computer Vision (ECCV) , 2024

    2024

  59. [59]

    ShareGPT-4o: Comprehensive Multimodal Annotations with GPT-4o

    OpenGVLab. ShareGPT-4o: Comprehensive Multimodal Annotations with GPT-4o. https:// sharegpt4o.github.io/, 2024

    2024

  60. [60]

    SVIT: Scaling up Visual Instruction Tuning

    Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. SVIT: Scaling up Visual Instruction Tuning. arXiv preprint arXiv:2307.04087, 2023

    Pith/arXiv arXiv 2023

  61. [61]

    Learning Transferable Visual Models from Natural Language Supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and Jack Clark et al. Learning Transferable Visual Models from Natural Language Supervision. In International Conference on Machine Learning (ICML), 2021

    2021

  62. [62]

    Implementation and Benchmarking of Perceptual Image Hash Functions (pHash)

    Christoph Zauner. Implementation and Benchmarking of Perceptual Image Hash Functions (pHash). Master’s thesis, Upper Austria University of Applied Sciences, 2010

    2010

  63. [63]

    Andrei Z. Broder. On the Resemblance and Containment of Documents. In Compression and Complexity of Sequences, 1997

    1997

  64. [64]

    SGDR: Stochastic Gradient Descent with Warm Restarts

    Ilya Loshchilov and Frank Hutter. SGDR: Stochastic Gradient Descent with Warm Restarts. In International Conference on Learning Representations (ICLR) , 2017

    2017

  65. [65]

    ScreenSpot-v2: A More Robust Screen UI Grounding Benchmark

    OS-Atlas Team. ScreenSpot-v2: A More Robust Screen UI Grounding Benchmark. arXiv preprint arXiv:2410.23218, 2024

    Pith/arXiv arXiv 2024

  66. [66]

    Li, and Yu Wu et al

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y .K. Li, and Yu Wu et al. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models (introduces GRPO). arXiv preprint arXiv:2402.03300, 2024. 34

    Pith/arXiv arXiv 2024

  67. [67]

    RLPR: Extrapolating RL VR to General Domains without Verifiers

    Tianyu Yu, Bo Ji, Shouli Wang, Shu Y ao, Zefan Wang, Ganqu Cui, Lifan Yuan, Ning Ding, Yuan Y ao, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. RLPR: Extrapolating RL VR to General Domains without Verifiers. arXiv preprint arXiv:2506.18254, 2025

    Pith/arXiv arXiv 2025

  68. [68]

    RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness

    Tianyu Yu, Haoye Zhang, Qiming Li, Qixin Xu, Yuan Y ao, Da Chen, Xiaoman Lu, Ganqu Cui, Yunkai Dang, Taiwen He, Xiaocheng Feng, Jun Song, Bo Zheng, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness. arXiv preprint arXiv:2405.17220, 2024

    arXiv 2024

  69. [69]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct Preference Optimization: Y our Language Model is Secretly a Reward Model. In Advances in Neural Information Processing Systems (NeurIPS) , 2023

    2023

  70. [70]

    Zico Kolter, and Zhuang Liu

    Mingjie Sun, Xinlei Chen, J. Zico Kolter, and Zhuang Liu. Massive Activations in Large Language Models. International Conference on Machine Learning (ICML) , 2024

    2024

  71. [71]

    Qualcomm AI Runtime (QAIRT) SDK.https://www.qualcomm.com/ developer/software/qualcomm-ai-engine-direct-sdk , 2024

    Qualcomm Technologies. Qualcomm AI Runtime (QAIRT) SDK.https://www.qualcomm.com/ developer/software/qualcomm-ai-engine-direct-sdk , 2024

    2024

  72. [72]

    Hexagon Tensor Processor (HTP): Qualcomm On-Device AI Accelerator

    Qualcomm Technologies. Hexagon Tensor Processor (HTP): Qualcomm On-Device AI Accelerator. https://www.qualcomm.com/products/technology/processors/hexagon, 2024

    2024

  73. [73]

    Qualcomm Genie: AI Runtime for On-Device Generative Models

    Qualcomm Technologies. Qualcomm Genie: AI Runtime for On-Device Generative Models. https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/ introduction.html, 2024

    2024

  74. [74]

    ONNX: Open Neural Network Exchange

    ONNX Community. ONNX: Open Neural Network Exchange. https://onnx.ai/, 2019

    2019

  75. [75]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Y e, Jie Shao, Zhaokai Wang, Zhe Chen, Hongjie Zhang, Ganlin Y ang, Haomin Wang, Qi Wei, Jinhui Yin, Wenhao Li, Erfei Cui, Guanzhou Chen, Zichen Ding, Changyao Tian, Zhenyu Wu, Jingjing Xie, Zehao Li, Bowen Y ang, Yuchen Duan, Xuehui Wang, Zhi H...

    Pith/arXiv arXiv 2025

  76. [76]

    Paech, Paul Pak, Rom N

    Alexander Amini, Anna Banaszak, Harold Benoit, Arthur Böök, Tarek Dakhran, Song Duong, Alfred Eng, Fernando Fernandes, Marc Härkönen, Anne Harrington, Ramin Hasani, Saniya Karwa, Yuri Khrustalev, Maxime Labonne, Mathias Lechner, Valentine Lechner, Simon Lee, 35 Zetian Li, Noel Loo, Jacob Marks, Edoardo Mosca, Samuel J. Paech, Paul Pak, Rom N. Parnichkun, ...

    arXiv 2025

  77. [77]

    OCRBench: On the Hidden Mystery of OCR in Large Multimodal Models

    Yuliang Liu, Zhang Li, Biao Y ang, Chunyuan Li, Xucheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. OCRBench: On the Hidden Mystery of OCR in Large Multimodal Models. Science China Information Sciences, 2024

    2024

  78. [78]

    OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning

    Ling Fu, Biao Y ang, Zhebin Kuang, Jiajun Song, Yuzhe Li, Linghao Zhu, Qidi Luo, Xinyu Wang, Hao Lu, and Mingxin Huang et al. OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning. arXiv preprint arXiv:2501.00321, 2025

    Pith/arXiv arXiv 2025

  79. [79]

    Berg, and Tamara L

    Licheng Yu, Patrick Poirson, Shan Y ang, Alexander C. Berg, and Tamara L. Berg. Modeling Context in Referring Expressions (RefCOCO). In European Conference on Computer Vision (ECCV), 2016

    2016

  80. [80]

    ࠞ”ཟb ശđ໡ᄝ “ࠞת

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient Memory Management for Large Language Model Serving with PagedAttention. In ACM Symposium on Operating Systems Principles (SOSP) , 2023. 36 A Qualitative Case Studies To intuitively present the actual performance of StepX-...

    2023