Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Guiding VLM Agents with Process Rewards at Inference Time for GUI Navigation

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read A process reward model that scores each candidate action at every GUI navigation step improves static action accuracy by 3.4% and raises dynamic task success by about a third.

desk verdict Useful static reranking result, but the dynamic gains are unverified and the paper's own numbers don't add up. read the letter →

arxiv 2504.16073 v1 pith:KXFFNXV4 submitted 2025-04-22 cs.CL

classification cs.CL
keywords processrewardmodelGUInavigationinference-timeguidanceactionrerankingvisuallanguageagentAndroid-in-the-WildOdysseyMind2Web
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 GuidNav, an inference-time method that uses a process reward model to choose which action a visual-language-model agent should take at each GUI navigation step. Instead of fine-tuning the acting model or waiting until the end of a trajectory to judge success, a separately trained reward model scores the agent's top candidate actions on the current screen, and the highest-scoring one is executed. The paper reports that this step-level reranking improves single-step action accuracy by an average of 3.4% across Android-in-the-Wild, GUI Odyssey, and Mind2Web static settings, and raises task success in a simulated dynamic Android environment from 43.7% to 58.1%, a roughly 33% relative gain over trajectory-level autonomous refinement. When reward-guided step selection is combined with trajectory reflection and retry, success reaches 71.6% after up to three retries. The interest of the claim is that it promises better GUI agents with black-box commercial policy models, without fine-tuning.

What carries the argument

The load-bearing object is the process reward model R(x, h_t, s_t, a_t), which maps the instruction, a language summary of the navigation history, the current screen, and a candidate action to a scalar reward. It is trained by mean squared error against step-level annotations, where the reward for an action is 1 if it matches ground truth and 0 otherwise under the dataset-specific matching rules. At inference it selects a* = argmax R(...) among k actions proposed by the policy VLM. The reward model's role is to reclaim the correct action that the policy already knows but does not rank first; the paper's oracle upper bound shows that gain is available in the candidate set.

What would settle it

Run the reward model on states collected from dynamic rollouts in the simulated Android environment: for each step, record the reward it assigns to the action that was actually executed, then compare with whether the trajectory eventually succeeds. If step-level reward accuracy on dynamic states is near chance, the static-trained reward model's transfer to dynamic environments is refuted, removing the basis for the headline 33% gain.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a process reward model trained on ground-truth-matched actions can serve as a reranker at inference time. Given a user instruction, a summary of prior steps, and the current screenshot, the policy VLM proposes k candidate next actions; the reward model assigns a scalar to each candidate in context, and the agent executes the argmax. Trained on human demonstrations plus VLM self-play trajectories, with rewards derived from matching the action type and target to ground truth (including a distance tolerance for clicks and direction matching for scrolls), the reward model achieves 78.8% step-wise accuracy on the AitW evaluation. The paper argues that this selection mechanism is what yields the gains: simply generating top-k actions does not help, but selecting among them with a learned reward does.

Load-bearing premise

The reward model is trained on static screenshots with labels derived from matching the predicted action to ground truth, and this matching proxy is assumed to transfer to dynamic environments where ground truth is absent and correctness depends on the evolving state.

Editorial extensions

If this is right

  • In static GUI benchmarks, reranking the policy's top-k actions with a learned process reward improves mean single-step action accuracy by 3.4% across Android-in-the-Wild, GUI Odyssey, and Mind2Web.
  • In a simulated dynamic Android setting, the same reward-guided selection raises task success to 58.1%, roughly 33% higher than the 43.7% trajectory-level autonomous-refinement baseline.
  • Combining process-reward selection with trajectory reflection and retry raises success further, to 71.6%, showing the two levels of supervision are complementary.
  • Because the acting VLM is only prompted, the recipe works with commercial black-box policy models and avoids expensive fine-tuning or online reinforcement learning.
  • The gap between reward-model selection and the oracle top-k upper bound indicates headroom: if candidate generation and reward scoring both improve, the ceiling is higher.

Reading between the lines

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

  • A testable extension the authors do not run is to separate two possible sources of gain: candidate quality (whether the correct action is in the top-k set) and reranking quality (whether the reward model finds it). Reporting both Pass@K and reward-selection accuracy on the same states would show how much of the 3.4% is due to each.
  • The dynamic result rests on an unvalidated transfer: the reward model is trained on static ground-truth matching, yet deployed where correctness depends on state evolution. A direct check would score the reward model on states encountered in dynamic rollouts and compare its rankings with whether the trajectory ultimately succeeds.
  • If the transfer holds, the same reranking recipe could be applied to any action space with cheap automatic labels, such as web navigation or desktop control, where the paper's own limitations section notes broader generalization is untested.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes GuidNav, an inference-time guidance method for GUI navigation agents. A process reward model, trained on human demonstrations and self-play data, assigns scalar rewards to k candidate actions generated by a VLM policy; the highest-reward action is executed at each step. Potential uses include integrating the reward model with trajectory-level reflection and retry. The authors evaluate in static settings on AitW, GUI Odyssey, and Mind2Web, and report a dynamic evaluation on 120 AitW tasks, comparing against DP, TopK, AR, and DigiRL. They report a 3.4% average static single-step action accuracy improvement and an approximately 33% relative dynamic task-success gain over the AR baseline (Table 3: GuidNav 58.1% vs AR 43.7%), with further gains from trajectory refinement.

Significance. GuidNav is a lightweight alternative to fine-tuning or online RL: it leaves the policy VLM untouched and requires only a compact process reward model. The static results across three benchmarks are directionally consistent, and the efficiency analysis in Table 6 addresses practical API cost. If reliable, the dynamic results would be an interesting demonstration that a small supervised reward model can improve task completion without policy updates. However, the dynamic claim is the most novel and currently the least supported component: the reward model is trained on a static ground-truth matching proxy and deployed in an unverifiable simulated setting, and the paper does not release code, data, or checkpoints. The internal numerical inconsistencies further reduce confidence in the reported effect sizes.

major comments (4)
  1. [Abstract vs. §5.1 and §7] The paper reports inconsistent headline dynamic gains. The Abstract states "around 33% increase in task success rate"; §5.1 states "overall improvements of approximately 17.8% and 14.4%" for GuidNav over DP and AR; and §7 concludes "around 15% increase in task success rate in dynamic settings." Given Table 3 (GuidNav 58.1, AR 43.7), the relative gain over AR is 14.4/43.7 = 32.9%, so the 15% figure in §7 is not explained. The static gain is also presented as 3.4% in the Abstract but "near 5%" in §7 without clarifying that the latter refers only to AitW. These headline numbers must be reconciled.
  2. [Tables 2 and 3] Several reported row averages are not the arithmetic means of the per-subset values. In Table 2, DP should be (30.3+39.1+36.2+34.4)/4 = 35.0, not 34.3; GuidNav should be (35.5+41.4+40.9+38.5)/4 = 39.1, not 38.9; and Pass@N should be (43.4+48.4+48.8+42.3)/4 = 45.7, not 46.8. In Table 3, DP should be 38.9, not 40.3; AR should be 41.9, not 43.7; GuidNav should be 56.4, not 58.1; and Integration should be 69.0, not 71.6. These discrepancies, up to 2.6 points, are comparable to or larger than several reported method-versus-baseline gaps, so the central comparative claims need to be recomputed and verified.
  3. [§4.2 and Table 3, dynamic evaluation] The dynamic evaluation is described only as "a simulated setting" with 120 sampled AitW tasks. The manuscript does not specify the simulator, the transition dynamics, how actions are executed, the task-success criterion used by the two human annotators, or inter-annotator agreement, and it provides no confidence intervals or significance tests. As a result, Table 3 is not interpretable as a reproducible benchmark result and the headline dynamic gain is currently unverifiable. The paper should report the full evaluation protocol and, ideally, release the code and simulator configuration.
  4. [§3.1, Eq. (3), Appendix A.1, §4.2] The process reward model is trained on static screenshots with labels derived from ground-truth-matching rules (Appendix A.1), then deployed in a dynamic environment where no ground truth exists and the correct action depends on the evolving state. The paper provides no evidence that PRM scores transfer to human-judged task success in the dynamic setting. A model trained on a static visual-matching proxy may systematically prefer actions that look like demonstrations but do not advance the task under changing environment dynamics. Please provide a direct validation, such as the correlation between PRM scores and eventual task completion on held-out dynamic episodes, or an offline analysis of whether the top-scoring actions would be allowed by the dynamic environment.
minor comments (6)
  1. [Abstract] The phrase "a around 33% increase" should be corrected to "an around 33% increase" or reworded.
  2. [Figure 2] The legend entry "TopK+AP" should read "TopK+AR" to match the text and the other abbreviations.
  3. [§6.3] The case study refers to "Figure ??", but the flight-search figure appears to be missing or its number is unresolved.
  4. [Appendix A.1] The phrase "For other actions: For other actions:" is duplicated and should be condensed.
  5. [Eq. (3)] The denominator is typeset as "1PN..." in a garbled way, and the notation switches between r^(i)_t,true and r^(i)_t,anno for the same quantity; please unify the notation and fix the equation formatting.
  6. [Table 3] The DigiRL row contains dashes for Google_apps and Install without explanation, and the statement that GuidNav "maintains superiority" over DigiRL is only partially supported because DigiRL is not evaluated on those subsets.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: static gains are supervised reranking on held-out tasks, and the only self-citation (TopK baseline) is non-load-bearing.

full rationale

The paper's central claim is empirical rather than a derivation: GuidNav trains a process reward model on labels derived from ground-truth action matching (Appendix A.1: 'a predicted action is considered correct if both the action type and the operation or gesture match the ground truth') and then uses that reward model to rerank candidate actions (Eqs. 5-6). The static evaluation metric is also defined as matching ground-truth actions ('we compute the screen-wise partial action matching score, where given the historical states, current state, and user instructions, the model's predicted action is compared against the ground-truth action', Section 4.3). This alignment is not circular: the reward model is fitted on a training split and evaluated on held-out tasks, so the reported static gains are a legitimate supervised-reranking result, not the training labels re-reported. The dynamic Android results (Table 3) rely on an unvalidated assumption that static matching scores transfer to the simulated dynamic environment, and no confidence intervals, code, or data are provided; that is a verifiability and correctness concern, not circularity under the definitions used here. The only self-citation is the TopK baseline description citing (Xiong et al., 2023; Tian et al., 2023), where Xiong et al. overlaps with the present authors. That citation is not load-bearing: TopK is a simple baseline ('TopK is a technique where the model generates the topk most probable actions'), and the paper's comparison does not rest on any substantive claim imported from the cited work. No equation reduces to its own input, no fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported via self-citation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the reward model's training proxy and its generalization. The listed free parameters (k=3, matching thresholds, trajectory count) are manually chosen and not ablated. The axioms are domain assumptions about proxy validity, self-play data quality, and simulator fidelity, none of which are independently verified.

free parameters (4)
  • k (number of action candidates) = 3
    The policy generates k candidate actions per step (Sections 3.2, 4.1); results likely depend on this choice, and no ablation is reported.
  • click matching distance threshold = 14% of screen distance
    Used to label reward model training data (Appendix A.1); manually chosen, affects training labels and therefore reward model behavior.
  • bounding box expansion factor = 240%
    Used in action matching for label assignment (Appendix A.1); manually chosen.
  • number of training trajectories per subset = 300
    Reward model training data scale; Table 1 shows performance varies with trajectory count (300 vs 800), so this choice affects results.
assumptions (4)
  • domain assumption The ground-truth matching rule used to label reward model training data is an adequate proxy for action correctness in GUI navigation.
    The reward model is trained to predict whether an action matches a human annotation under the Appendix A.1 rules; if this proxy is wrong, the model selects actions that match the rule but not the true intent.
  • domain assumption The process reward model trained on static, ground-truth-annotated states generalizes to dynamic environments without ground truth.
    Dynamic evaluation (Section 4.2, Table 3) relies on this transfer; no evidence is provided that step-level correctness is stable across state transitions.
  • domain assumption GPT-4o-generated self-play trajectories provide valid training signal for the reward model.
    Self-play data (Section 3.1, Appendix A.1) is labeled by matching to ground truth, but the set of states and actions explored by GPT-4o may not cover the diversity of the deployment distribution.
  • domain assumption The simulated dynamic environment faithfully reproduces real Android GUI behavior.
    Dynamic results are collected in a 'simulated setting' (Section 4.2); the simulator is never described, so the external validity of the reported success rates is unclear.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Guiding VLM Agents with Process Rewards at Inference Time for GUI Navigation." pith.science (2026). https://pith.science/paper/KXFFNXV4

@misc{pith2026250416073,
  author       = {Pith},
  title        = {Pith review of: Guiding VLM Agents with Process Rewards at Inference Time for GUI Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXFFNXV4}},
  note         = {Machine review of arXiv:2504.16073}
}
read the original abstract

Recent advancements in visual language models (VLMs) have notably enhanced their capabilities in handling complex Graphical User Interface (GUI) interaction tasks. Despite these improvements, current frameworks often struggle to generate correct actions in challenging GUI environments. State-of-the-art commercial VLMs are black-boxes, and fine-tuning open-source VLMs for GUI tasks requires significant resources. Additionally, existing trajectory-level evaluation and refinement techniques frequently fall short due to delayed feedback and local optimization issues. To address these challenges, we propose an approach that guides VLM agents with process supervision by a reward model during GUI navigation and control at inference time. This guidance allows the VLM agent to optimize actions at each inference step, thereby improving performance in both static and dynamic environments. In particular, our method demonstrates significant performance gains in three GUI navigation tasks, achieving a 3.4% improvement in single step action accuracy for static environments, along with a around 33% increase in task success rate in one dynamic environment. With further integration of trajectory reflection and retry mechanisms, we also demonstrate even greater enhancement in task success.

Figures

Figures reproduced from arXiv: 2504.16073 by the authors.

Figure 1
Figure 1. Overview of GuidNav. optimization may degrade overall performance. Moreover, trajectory-level evaluation delays error correction, increasing both computational and time costs. Meanwhile, Bai et al. (2024) propose DigiRL, improving task performance in dynamic environments by combining Advantage-Weighted Regression with online reinforcement learning (RL) and an automatic curriculum mechanism. Such RL methods can lead … view at source ↗
Figure 2
Figure 2. The performance curve across different trial numbers shows the impact of refinement [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Example of case study. Access the accessibility settings. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID 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

  1. Software Engineering for and with GUI Agent

    cs.SE 2026-08 conditional novelty 5.0 of 10

    A survey of 336 GUI-agent papers finds rapid growth alongside weak engineering support for recovery, human oversight, maintainability, and privacy, and calls for lifecycle-centered testing and governance.

Reference graph

Works this paper leans on

27 extracted references · 5 canonical work pages · cited by 1 Pith paper

  1. [3]

    arXiv preprint arXiv:2401.10935 (2024)

    Seeclick: Harnessing GUI grounding for advanced visual GUI agents. arXiv preprint arXiv:2401.10935 (2024). Google DeepMind

  2. [4]

    https://deepmind.google/technologies/ gemini/flash/ Accessed: 2025-02-15

    Gemini 2.0 Flash. https://deepmind.google/technologies/ gemini/flash/ Accessed: 2025-02-15. Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su

  3. [5]

    Advances in Neural Information Processing Systems 36 (2023), 28091–28114

    Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems 36 (2023), 28091–28114. Moghis Fereidouni and AB Siddique

  4. [6]

    arXiv preprint arXiv:2404.10887 (2024)

    Search Beyond Queries: Training Smaller Language Models for Web Interactions via Reinforcement Learning. arXiv preprint arXiv:2404.10887 (2024). Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. 2024a. CogVLM2: Visual Language Models for Image and Video Understanding. arXiv preprint...

  5. [7]

    arXiv preprint arXiv:2410.21276 (2024)

    Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024). Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  6. [8]

    arXiv preprint arXiv:2401.13649 (2024)

    Visualwebarena: Evaluating multimodal agents on realistic visual web tasks. arXiv preprint arXiv:2401.13649 (2024). Quanfeng Lu, Wenqi Shao, Zitao Liu, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, Yu Qiao, and Ping Luo

  7. [9]

    arXiv preprint arXiv:2406.08451 (2024)

    Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. arXiv preprint arXiv:2406.08451 (2024). Xing Han Lù, Zdenˇek Kasner, and Siva Reddy

  8. [10]

    arXiv preprint arXiv:2402.05930 (2024)

    Weblinx: Real-world website navigation with multi-turn dialogue. arXiv preprint arXiv:2402.05930 (2024). OpenAI

Show all 27 references
  1. [11]

    arXiv preprint arXiv:2303.08774 (2023)

    GPT-4 Technical Report. arXiv preprint arXiv:2303.08774 (2023). OpenAI

  2. [12]

    arXiv preprint arXiv:2404.06474 (2024)

    Au- tonomous evaluation and refinement of digital agents. arXiv preprint arXiv:2404.06474 (2024). Christopher Rawles, Sarah Clinckemaillie, Yifan Chang, Jonathan Waltz, Gabrielle Lau, Marybeth Fair, Alice Li, William Bishop, Wei Li, Folawiyo Campbell-Ajala, et al. 2024a. Andro...

  3. [13]

    arXiv preprint arXiv:2403.05530 (2024)

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 (2024). Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao

  4. [14]

    Advances in Neural Information Processing Systems 36 (2024)

    Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems 36 (2024). Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning

  5. [15]

    arXiv preprint arXiv:2305.14975 (2023)

    Just ask for calibration: Strategies for eliciting cali- brated confidence scores from language models fine-tuned with human feedback. arXiv preprint arXiv:2305.14975 (2023). Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang

  6. [16]

    arXiv preprint arXiv:2406.01014 (2024)

    Mobile-Agent-v2: Mobile Device Operation Assistant with Effective Navigation via Multi-Agent Collaboration. arXiv preprint arXiv:2406.01014 (2024). Zhiyong Wu, Chengcheng Han, Zichen Ding, Zhenmin Weng, Zhoumianze Liu, Shunyu Yao, Tao Yu, and Lingpeng Kong

  7. [17]

    arXiv preprint arXiv:2402.07456 (2024)

    OS-copilot: Towards generalist computer agents with self-improvement. arXiv preprint arXiv:2402.07456 (2024). Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al

  8. [18]

    arXiv preprint arXiv:2404.07972 (2024)

    OSworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972 (2024). Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi

  9. [19]

    arXiv preprint arXiv:2306.13063 (2023)

    Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. arXiv preprint arXiv:2306.13063 (2023). An Yan, Zhengyuan Yang, Wanrong Zhu, Kevin Lin, Linjie Li, Jianfeng Wang, Jianwei Yang, Yiwu Zhong, Julian McAuley, Jianfeng Gao, et al

  10. [20]

    arXiv preprint arXiv:2311.07562 (2023)

    Gpt-4v in wonderland: Large multimodal models for zero-shot smartphone GUI navigation. arXiv preprint arXiv:2311.07562 (2023). Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. 2023b. Set-of-mark prompting unleashes extraordinary visual grounding in ...

  11. [21]

    arXiv preprint arXiv:2404.05719 (2024)

    Ferret-UI: Grounded Mobile UI Understanding with Multimodal LLMs. arXiv preprint arXiv:2404.05719 (2024). Yuexiang Zhai, Hao Bai, Zipeng Lin, Jiayi Pan, Shengbang Tong, Yifei Zhou, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma, et al

  12. [22]

    arXiv preprint arXiv:2405.10292 (2024)

    Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning. arXiv preprint arXiv:2405.10292 (2024). Zhuosheng Zhan and Aston Zhang

  13. [23]

    arXiv preprint arXiv:2309.11436 (2023)

    You only look at screens: Multimodal chain-of-action agents. arXiv preprint arXiv:2309.11436 (2023). Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang

  14. [24]

    arXiv preprint arXiv:2403.02713 (2024)

    Android in the zoo: Chain-of-action-thought for GUI agents. arXiv preprint arXiv:2403.02713 (2024). Longtao Zheng, Zhiyuan Huang, Zhenghai Xue, Xinrun Wang, Bo An, and Shuicheng Yan

  15. [25]

    arXiv preprint arXiv:2403.17918 (2024)

    Agentstudio: A toolkit for building general virtual agents. arXiv preprint arXiv:2403.17918 (2024). 13 Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee

  16. [26]

    Advances in Neural Information Processing Systems 36 (2024)

    Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024). 14 A Appendix A.1 Reward Annotation To collect training data for the reward model, we utilize three datasets: AitW (Rawles et al., 2024b), GUI Odyssey (Lu et al., 2024), and...

  17. [27]

    Each region is annotated with marks like alphanumeric labels, masks, or bounding boxes

    or SAM (Kirillov et al., 2023), to partition an image into regions of varying granularity. Each region is annotated with marks like alphanumeric labels, masks, or bounding boxes. This enhances the VLM’s ability to interpret and understand image elements. In our implementation,...

  18. [2023]

    arXiv preprint arXiv:2308.12966 (2023)

    Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilities. arXiv preprint arXiv:2308.12966 (2023). Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu

  19. [2024]

    arXiv preprint arXiv:2406.11896 (2024)

    Digirl: Training in-the-wild device-control agents with autonomous reinforcement learning. arXiv preprint arXiv:2406.11896 (2024). Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.