The Next Screenshot Knows: Gated Hindsight Distillation for Mobile GUI Agents

Hengfu Yu; Junzhuo Liu; Tong Chu; Weiwei Li; Wen Li

REVIEW 2 major objections 6 minor 18 references

The Next Screenshot Knows: Gated Hindsight Distillation for Mobile GUI Agents

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A training-time teacher that sees the next screenshot improves mobile GUI agent task success.

desk verdict Solid empirical paper with a real, consistent gain for GUI-agent training, but the 'next screenshot knows' mechanism is partially confounded by the teacher's access to the student's own reasoning tokens. read the letter →

arxiv 2608.06065 v1 pith:VQRUNN43 submitted 2026-08-06 cs.CV

Weiwei Li , Junzhuo Liu , Tong Chu , Hengfu Yu , Wen Li This is my paper · ORCID

classification cs.CV

keywords GUIagentsmobilehindsightdistillationprivilegedinformationreinforcementlearningGRPOvision-languagemodelofflinetraining

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

Offline training of GUI agents normally throws away the next screenshot after each demonstrated action, even though that screenshot often contains the only evidence that justifies the action. Gated Hindsight Distillation (GHD) instead treats the next screenshot as privileged training information: a parameter-sharing teacher sees it while re-scoring the student's rollouts, and the corrected token-level distribution is distilled into the student only when the student fails and the teacher's recovery matches the demonstrated action. The paper claims this future-grounded supervision improves task success over SFT and GRPO on AndroidWorld and AndroidLab with both Qwen2.5-VL-7B and Qwen3-VL-8B. If true, GHD offers a training-only way to inject application-specific GUI knowledge into a prefix-only policy, with no extra inference-time module.

What carries the argument

The load-bearing mechanism is Gated Hindsight Distillation: a joint objective $L = L_{\text{GRPO}} + \lambda L_{\text{GHD}}$ with $\lambda = 0.1$, where $L_{\text{GHD}}$ is a generalized Jensen–Shannon divergence ($\alpha = 0.5$) between the stop-gradient teacher distribution $\pi_T(\cdot \mid \tilde{x}_t, y_{<j})$ and the trainable student distribution $\pi_S(\cdot \mid x_t, y_{<j})$, evaluated on the student's top-$K = 100$ tokens plus a residual bucket. The gate $M(y)$ keeps a response only when the step reward is below $\tau_{\text{succ}} = 1.45$ (student fails) and the teacher's position-wise top-1 reconstruction matches the demonstrated action within tolerance $\delta = 20$ for coordinates or exact/edit similarity for text. Dynamic sampling draws up to three rollout-group attempts per prompt to increase the density of gate-passing responses. Because the teacher shares parameters with the student and differs only in receiving the next observation, the entire method adds no inference-time module.

What would settle it

Replace the next screenshot with a randomly chosen screenshot from a different task while keeping the gate and distillation unchanged; if Pass@1 does not fall, the specific content of the next screenshot is not providing the corrective signal, falsifying the claimed hindsight mechanism.

Watch

Extended reading notes

Core claim

The paper's central claim is that the next screenshot, observed only during training, converts a hard prediction problem ('what should I do?') into an easier inference problem ('what must I have done?') and that the resulting rationale can be distilled into a prefix-only student. GHD implements this with a parameter-sharing teacher that conditions on the student's own rollout tokens plus the realized next observation, computes teacher-forced token distributions, and applies a gated distillation loss that keeps only the cases where the prefix-only student fails and the teacher's position-wise top-1 action matches the demonstration. The reported results on AndroidWorld and AndroidLab show consistent Pass@1 gains over SFT and GRPO across two vision-language model scales, with the biggest improvements on actions that require application-specific navigation knowledge or implicit prerequisites.

Load-bearing premise

The method's claimed mechanism—that the future screenshot is the source of the correction—is not isolated from the student's own generated reasoning tokens, which the teacher also sees and which could already contain the demonstrated action.

Editorial extensions

If this is right

GHD improves task success (Pass@1) over both SFT and GRPO on AndroidWorld and AndroidLab for Qwen2.5-VL-7B and Qwen3-VL-8B, and the gains are largest on steps that require application-specific navigation knowledge or implicit prerequisites.
The step-wise gains come mainly from the future observation itself: ablations show it contributes the largest increment over gating and dynamic sampling, and it outperforms providing the ground-truth action or reasoning alone.
Continuous token-level distribution matching transfers privileged knowledge better than STaR-style off-policy rationalization for every privileged signal tested.
Compared with a GUI-Shift-style inverse-dynamics auxiliary task, GHD's direct future-grounded distillation yields a much larger improvement, indicating that the transfer mechanism matters, not just the future state.
At deployment GHD removes the teacher, the future observation, and the gating procedure, so the prefix-only student runs with no additional inference cost.

Reading between the lines

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

A testable extension would isolate the contribution of the future screenshot by removing the student's own reasoning tokens ($y_{<j}$) from the teacher's input; if the gate passes purely on those tokens, the reported gains could be due to self-hints rather than hindsight.
The same gated hindsight scheme could be applied to other sequential decision tasks—web navigation, form filling, robotics—where a verifiable step reward and demonstration trajectories are available and later observations reveal action consequences.
A soft gate that weights distillation by the teacher's confidence or the margin of correction might capture more signal than the hard threshold at $\tau_{\text{succ}} = 1.45$, and could reduce the need for the 2.69-attempts-per-batch dynamic sampling overhead.
The per-application breakdown (gains on seven of nine AndroidLab apps, none on Calendar and Zoom) suggests the method's benefit is concentrated in transition types that the gate selects, so a future analysis could predict where hindsight distillation will help from the gate's acceptance statistics.

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

Short take: GHD is a real, reproducible-looking gain for offline GUI-agent training, and the experiments are better controlled than most papers in this area. The main caveat is that the teacher conditions on the student's own rollout prefix, so the 'next screenshot knows' framing is stronger than the evidence.

What's good: The method combines existing pieces—future-conditioned hindsight (as in GUI-Shift, but used only at training), on-policy self-distillation (SDPO), and a gating rule that only distills when the student fails and the teacher verifiably recovers the demonstrated action. The gating and dynamic sampling are simple and clearly ablated. Tables 2 and 3 are the core: component-wise ablation on AndroidWorld with 7B, and controlled comparison with means/stds over three runs on both benchmarks and two model families. The gains hold up: 5-11 points over GRPO. Table 4 shows that adding the next screenshot gives a larger jump than adding the reference action or reasoning, which supports the idea that the future observation carries useful signal. The comparison with a GUI-Shift-style auxiliary task is also fair and shows GHD's more direct supervision works better.

Soft spots: The main one is the teacher's conditioning on y

Who this is for: anyone working on offline training of GUI agents, especially RL/distillation pipelines. It deserves a serious referee. The right request would be: add the y

Referee Report

2 major / 6 minor

Summary. The paper proposes Gated Hindsight Distillation (GHD), a training-time method for mobile GUI agents. A student policy generates a reasoning-and-action response from the observable interaction prefix, while a parameter-sharing teacher additionally sees the next screenshot from the successful demonstration trajectory and re-scores the student's sampled tokens. Distillation is gated to rollouts where the prefix-only student fails and the teacher's position-wise top-1 decoding recovers the demonstrated action, and the objective is combined with GRPO. Experiments on AndroidWorld and AndroidLab with Qwen2.5-VL-7B and Qwen3-VL-8B report consistent Pass@1 gains over SFT and GRPO, with ablations for gating, dynamic sampling, and the privileged next-screenshot observation.

Significance. If the reported results hold, GHD is a practical and simple way to exploit future observations that are already present in offline successful trajectories, addressing a genuine supervision gap in GUI-agent training without adding any inference-time component. The paper's strengths are its controlled comparisons with means and standard deviations over three runs, a clean ablation structure that separates gating, dynamic sampling, and the privileged next screenshot, and direct comparisons against alternative uses of future states such as STaR-style target generation and a GUI-Shift-style inverse-dynamics auxiliary task. The method is described in enough detail to be reproducible in principle, and the code/checkpoint availability statement is a positive commitment, though it should be more concrete.

major comments (2)

[Experiments / GHD Implementation Details] Please clarify the training data for the SFT baseline. The 'GHD Implementation Details' paragraph states that all experiments use the same SFT checkpoint, training split, rollout configuration, and environment for SFT, GRPO, and GHD, but the following sentence says the training data is a hard subset of the OpenMobile trajectories obtained by filtering examples that the corresponding SFT model solves in one attempt. If the SFT row in Table 3 is the checkpoint trained on the full 27,360 examples while GRPO and GHD are trained on the 6,968 (or 5,982) hard examples, then the 'Controlled Comparison' section's claim of 'the same data' is not valid for the SFT row. If the SFT baseline was retrained on the hard subset, state that explicitly. This is needed to interpret the SFT-to-GRPO and SFT-to-GHD comparisons in Table 3.
[Ablation Studies (Table 2)] The component-wise ablation in Table 2 is reported as single-run numbers. The text claims that the next screenshot provides the largest incremental gain (+3.17 over +DS) and that 'the majority of the improvement comes from transferring future-grounded token-level supervision,' but no variance is reported for these ablation conditions. Given that Table 3 shows run-to-run standard deviations of 0.65-1.51 on the same benchmark, the 0.71, 2.43, and 3.17 point differences could be within noise. Please report means and standard deviations over at least three runs for the ablation conditions, or explicitly temper the mechanistic claim.

minor comments (6)

[Method, Eq. (5); Ablation Studies (Table 2)] Equation (5) conditions the teacher on the student's sampled prefix y<j; the paper should state explicitly that the +DS row in Table 2 is the control for this choice, because it keeps y<j while removing only ot+1. Such a statement would preempt the alternative explanation that the student's own reasoning tokens drive the correction signal.
[Privileged Information and Transfer Mechanism (Table 4)] In Table 4, the 'Ours' row (only ot+1) outperforms the 'Full' row (a*, r*, ot+1) by 1.80 points; please discuss why adding the ground-truth action and reasoning hurts performance.
[Efficiency and Sampling Overhead] In the sentence 'GHD without dynamic sampling still improves over GRPO by by 2.29 points on AndroidWorld,' the word 'by' is duplicated.
[Experiments, Table 7] Table 7 reports a single representative run; please indicate whether the per-application numbers are averaged over the three runs and, if not, state that the breakdown is illustrative.
[Abstract and Code Availability] The abstract and experiments state that code and checkpoints 'will be made available'; for reproducibility, please provide a link or a concrete release timeline in the submission.
[Experiments / GHD Implementation Details] Please describe the hard-subset filtering step in more detail: which model performed the filtering, how 'solves in one attempt' is determined with the step verifier, and whether the SFT baseline in Table 3 was trained before or after filtering.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GHD is an empirical training method whose objective is defined, not fitted, and whose central future-screenshot claim is supported by controlled ablations that hold the student prefix fixed.

full rationale

The paper does not contain a derivation chain whose output is equivalent to its input by construction. The central loss, L = L_GRPO + λL_GHD (Eq. 10), is a defined training objective with hand-set hyperparameters (λ=0.1, τ_succ=1.45, δ=20, K=100), not a fitted equation or a self-consistency condition. The teacher is parameter-sharing and conditioned on the student's sampled prefix y<j and the next screenshot ot+1 (Eq. 5), but this is a transparent design choice, not a circular reduction: the claim that ot+1 is the primary source of improvement is tested by ablations that keep y<j present while removing future information. Table 2 compares GRPO, +Gate (an unprivileged teacher with the same student-prefix conditioning), +DS, and full GHD, and the largest incremental gain appears when the next screenshot is added, so the future observation is not assumed into the result by definition. Table 4 further varies the privileged information (action, reasoning, next observation) while holding gating, dynamic sampling, rollout, and optimization settings fixed. The benchmark comparisons on AndroidWorld and AndroidLab are external and independent of any fitted quantity. Self-citation is not load-bearing: the SDPO reference supplies only the Jensen-Shannon divergence formulation and is external to this paper's authors, and no uniqueness theorem or ansatz is imported from prior work by the same authors. The concern that the student's own prefix may already state the correct action is an empirical confounding possibility rather than a definitional circularity, and it is partly addressed by the unprivileged-teacher controls. The evaluation is self-contained against external benchmarks, so the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

GHD is an empirical method paper and introduces no new physical or mathematical entities. The central claim rests on domain assumptions about future screenshots carrying rationale, greedy teacher decoding matching distillation behavior, and verifier reliability, plus the hand-selected hyperparameters listed above. The parameter-sharing 'teacher' is a model modulation, not a new entity, and no equation in the paper is a derivation of a predicted constant from fitted parameters.

free parameters (5)

distillation weight lambda = 0.1
Hand-selected weight for the GHD loss in Eq. 10; controls how much the future-grounded distillation term contributes relative to GRPO.
gate threshold tau_succ = 1.45
Step reward threshold above which a response is considered successful and excluded from distillation; hand-selected in GHD Implementation Details.
coordinate tolerance delta = 20 on normalized grid
Tolerance for matching click, long_press, and swipe coordinates in the gate condition of Eq. 9.
top-K vocabulary K = 100
Number of student tokens kept when approximating the Jensen-Shannon divergence in Eq. 6; fixed in all experiments.
symmetric divergence weight alpha = 0.5
Balances the two KL terms in the generalized Jensen-Shannon divergence of Eq. 6; hand-selected and fixed.

assumptions (3)

domain assumption The next screenshot from a successful trajectory reliably reveals the rationale for the demonstrated action.
Used throughout the Motivation and Gating sections to argue that future observation supplies the missing evidence; the paper does not directly validate this premise beyond anecdotal examples.
domain assumption Position-wise top-1 decoding of the teacher distribution is an accurate proxy for the teacher's corrected action and for the signal used in distillation.
The gate in Eq. 8 selects episodes based on greedy per-token argmax decoding; this may not match full rollout behavior and is not evaluated as a proxy.
domain assumption The step verifier reward R(y) in Eq. 3 correctly identifies successful and failed responses for both GRPO training and gating.
All training and gating decisions depend on this scalar reward, but the paper reports no validation of verifier quality or comparison to human judgment.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Next Screenshot Knows: Gated Hindsight Distillation for Mobile GUI Agents." pith.science (2026). https://pith.science/paper/VQRUNN43

@misc{pith2026260806065,
  author       = {Pith},
  title        = {Pith review of: The Next Screenshot Knows: Gated Hindsight Distillation for Mobile GUI Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VQRUNN43}},
  note         = {Machine review of arXiv:2608.06065}
}

read the original abstract

GUI agents are commonly trained offline from successful interaction trajectories. Standard training decomposes each trajectory into prefix-action pairs: the agent predicts an action from the current screen and interaction history, while the subsequent observation is discarded. This removes the rationale of why an action is correct: the evidence often appears only on the subsequent screen. For example, to enable Soft Wrap, the agent should click Edit or View, but nothing reveals this until the menu opens. Without such evidence, standard imitation gives the model little chance of ever sampling and thus learning the correct reasoning. To address this issue, we propose Gated Hindsight Distillation (GHD), which uses the next screenshot as privileged information during training. A student predicts from the observable trajectory prefix, while a parameter-sharing teacher additionally observes the next screenshot and re-scores the student's on-policy responses. We apply distillation only when the student fails and the hindsight-conditioned teacher recovers the demonstrated action. GHD improves task success over GRPO on AndroidWorld and AndroidLab across two vision-language models. The code and checkpoints will be made available.

Figures

Figures reproduced from arXiv: 2608.06065 by the authors.

**Figure 2.** Overview of Gated Hindsight Distillation. Given a successful trajectory, the student observes the interaction prefix [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Effect of privileged information and transfer mech [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 4 canonical work pages

[1]

Qwen3-vl technical report.arXiv preprint arXiv:2511.21631

Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng,L.;Ding,W.;Gao,C.;Ge,C.;etal.2025a. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631. Bai,S.;Chen,K.;Liu,X.;Wang,J.;Ge,W.;Song,S.;Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y.; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y.; Ye, J.; Zhang, X.; Xie, T.; Che...

arXiv
[4]

Guan,Y.;Yu,R.;Zhang,J.;Wang,L.;Zhang,C.;Li,L.;Qiao, B.;Qin,S.;Huang,H.;Yang,F.;etal.2026

Ui-venus technical report: Building high-performance ui agents with rft.arXiv preprint arXiv:2508.10833. Guan,Y.;Yu,R.;Zhang,J.;Wang,L.;Zhang,C.;Li,L.;Qiao, B.;Qin,S.;Huang,H.;Yang,F.;etal.2026. Computer-using world model.arXiv preprint arXiv:2602.17365. Guo, D.; Yang, D.; Zhang, H.; Song, J.; Wang, P.; Zhu, Q.; Xu, R.; Zhang, R.; Ma, S.; Bi, X.; et al

arXiv 2026
[5]

Huang, J.; Huang, Z.; Shi, Y.; Yang, T.; Zhai, X.; Chu, W.; and Liu, N

GUI Agents with Reinforce- ment Learning: Toward Digital Inhabitants.arXiv preprint arXiv:2604.27955. Huang, J.; Huang, Z.; Shi, Y.; Yang, T.; Zhai, X.; Chu, W.; and Liu, N

arXiv
[6]

Trust the Right Teacher: Quality-Aware Self-Distillation for GUI Grounding

Trust the Right Teacher: Quality- Aware Self-Distillation for GUI Grounding.arXiv preprint arXiv:2606.18101. Hübotter,J.;Lübeck,F.;Behric,L.;Baumann,A.;Bagatella, M.; Marta, D.; Hakimi, I.; Shenfeld, I.; Buening, T. K.; Guestrin, C.; et al

work page Pith review arXiv
[7]

Im, Y.; Jo, B.; Wi, J.; Baek, S.; Min, T

Reinforcement Learning via Self- Distillation.arXiv preprint arXiv:2601.20802. Im, Y.; Jo, B.; Wi, J.; Baek, S.; Min, T. H.; Lee, J. H.; Oh, S.; Shin, I.; and Lee, S

arXiv
[8]

Li, W.; Bishop, W.; Li, A.; Rawles, C.; Campbell-Ajala, F.; Tyamagundlu, D.; and Riva, O

Modular and Multi-Path- AwareOfflineBenchmarkingforMobileGUIAgents.arXiv preprint arXiv:2512.12634. Li, W.; Bishop, W.; Li, A.; Rawles, C.; Campbell-Ajala, F.; Tyamagundlu, D.; and Riva, O

arXiv
[9]

Liu,Y.;Liu,Z.;Zhu,S.;Li,P.;Xie,C.;Wang,J.;Hu,X.;Han, X.; Yuan, J.; Wang, X.; et al

Ui-voyager: A self-evolving gui agent learning via failed experience.arXiv preprint arXiv:2603.24533. Liu,Y.;Liu,Z.;Zhu,S.;Li,P.;Xie,C.;Wang,J.;Hu,X.;Han, X.; Yuan, J.; Wang, X.; et al

arXiv
[10]

arXiv preprint arXiv:2509.15221

Scalecua: Scaling open-source computer use agents with cross-platform data. arXiv preprint arXiv:2509.15221. Lu, Q.; Shao, W.; Liu, Z.; Du, L.; Meng, F.; Li, B.; Chen, B.;Huang,S.;Zhang,K.;andLuo,P.2025a. Guiodyssey:A comprehensive dataset for cross-app gui navigation on mo- bile devices. InProceedings of the IEEE/CVF International Conference on Computer ...

arXiv

Show all 18 references

[11]

Lu, Z.; Ye, J.; Tang, F.; Shen, Y.; Xu, H.; Zheng, Z.; Lu, W.; Yan, M.; Huang, F.; Xiao, J.; et al

Ui-r1: Enhancing efficient action prediction of gui agents by reinforcement learning.InProceedingsoftheAAAIConferenceonArtificial Intelligence, volume 40, 17608–17616. Lu, Z.; Ye, J.; Tang, F.; Shen, Y.; Xu, H.; Zheng, Z.; Lu, W.; Yan, M.; Huang, F.; Xiao, J.; et al. 2025b. Ui...
[12]

Rawles, C.; Clinckemaillie, S.; Chang, Y.; Waltz, J.; Lau, G.; Fair, M.; Li, A.; Bishop, W.; Li, W.; Campbell-Ajala, F.; et al

Ui-tars: Pioneering automated gui interaction with native agents.arXiv preprint arXiv:2501.12326. Rawles, C.; Clinckemaillie, S.; Chang, Y.; Waltz, J.; Lau, G.; Fair, M.; Li, A.; Bishop, W.; Li, W.; Campbell-Ajala, F.; et al

arXiv
[13]

InInternationalConference on Learning Representations, volume 2025, 406–441

Androidworld: A dynamic benchmarking envi- ronmentforautonomousagents. InInternationalConference on Learning Representations, volume 2025, 406–441. Shao,R.;Gao,R.;Xie,B.;Li,Y.;Zhou,K.;Wang,S.;Guan, W.; and Chen, G

2025
[14]

InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 27471–27481

Hats: Hardness-aware trajectory synthesis for gui agents. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 27471–27481. Team,V.;Gao,C.;Gu,Z.;Liu,Y.;Qiu,X.;Shen,S.;Wen,Y.; Xia,T.;Xu,Z.;Zeng,Z.;etal.2026. Ui-venus-1.5technical report.arXiv pre...

2026
[16]

Zelikman,E.;Wu,Y.;Mu,J.;andGoodman,N.2022

Step-gui technical report.arXiv preprint arXiv:2512.15431. Zelikman,E.;Wu,Y.;Mu,J.;andGoodman,N.2022. STaR: Bootstrapping Reasoning With Reasoning. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A., eds.,Advances in Neural Information Processing Sys- t...

2022
[17]

InProceedings of the 2025 Conference on Empirical Meth- ods in Natural Language Processing: System Demonstra- tions, 155–180

Agentcpm-gui: Building mobile-use agents with reinforcement fine-tuning. InProceedings of the 2025 Conference on Empirical Meth- ods in Natural Language Processing: System Demonstra- tions, 155–180. Zhou, H.; Zhang, X.; Tong, P.; Zhang, J.; Chen, L.; Kong, Q.; Cai, C.; Liu, C....

2025
[18]

MAI- UI Technical Report: Real-World Centric Foundation GUI Agents.arXiv preprint arXiv:2512.22047
[2024]

Wu, D.; Hao, R.; Wang, H.; Wu, S.; Xiao, H.; Li, Z.; Zhou, B.; Ju, Z.; Liu, Z.; Fan, L.; et al

Gui agents with foundation models: A comprehensive survey.arXiv preprint arXiv:2411.04890. Wu, D.; Hao, R.; Wang, H.; Wu, S.; Xiao, H.; Li, Z.; Zhou, B.; Ju, Z.; Liu, Z.; Fan, L.; et al. 2026a. Mobilegym: A verifiableandhighlyparallelsimulationplatformformobile gui agent resea...

arXiv
[2025]

InFindings oftheAssociationforComputationalLinguistics:ACL2025, 2138–2156

Amex: Android multi- annotation expo dataset for mobile gui agents. InFindings oftheAssociationforComputationalLinguistics:ACL2025, 2138–2156. Cheng,K.;Li,Z.;Ma,Z.;Chen,N.;Cao,J.;Sun,Q.;Ding,Z.; Xu,F.;Yan,H.;Chen,J.;etal.2026. OpenMobile:Building openmobileagentswithtaskandtra...

2026 arXiv
[2026]

Gao, L.; Zhang, L.; Gao, P.; Liu, W.; Luan, J.; and Xu, M

Skill-Guided Continuation Distillation for GUI Agents.arXiv preprint arXiv:2606.18890. Gao, L.; Zhang, L.; Gao, P.; Liu, W.; Luan, J.; and Xu, M

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Qwen3-vl technical report.arXiv preprint arXiv:2511.21631

Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng,L.;Ding,W.;Gao,C.;Ge,C.;etal.2025a. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631. Bai,S.;Chen,K.;Liu,X.;Wang,J.;Ge,W.;Song,S.;Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y.; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y.; Ye, J.; Zhang, X.; Xie, T.; Che...

arXiv

[2] [4]

Guan,Y.;Yu,R.;Zhang,J.;Wang,L.;Zhang,C.;Li,L.;Qiao, B.;Qin,S.;Huang,H.;Yang,F.;etal.2026

Ui-venus technical report: Building high-performance ui agents with rft.arXiv preprint arXiv:2508.10833. Guan,Y.;Yu,R.;Zhang,J.;Wang,L.;Zhang,C.;Li,L.;Qiao, B.;Qin,S.;Huang,H.;Yang,F.;etal.2026. Computer-using world model.arXiv preprint arXiv:2602.17365. Guo, D.; Yang, D.; Zhang, H.; Song, J.; Wang, P.; Zhu, Q.; Xu, R.; Zhang, R.; Ma, S.; Bi, X.; et al

arXiv 2026

[3] [5]

Huang, J.; Huang, Z.; Shi, Y.; Yang, T.; Zhai, X.; Chu, W.; and Liu, N

GUI Agents with Reinforce- ment Learning: Toward Digital Inhabitants.arXiv preprint arXiv:2604.27955. Huang, J.; Huang, Z.; Shi, Y.; Yang, T.; Zhai, X.; Chu, W.; and Liu, N

arXiv

[4] [6]

Trust the Right Teacher: Quality-Aware Self-Distillation for GUI Grounding

Trust the Right Teacher: Quality- Aware Self-Distillation for GUI Grounding.arXiv preprint arXiv:2606.18101. Hübotter,J.;Lübeck,F.;Behric,L.;Baumann,A.;Bagatella, M.; Marta, D.; Hakimi, I.; Shenfeld, I.; Buening, T. K.; Guestrin, C.; et al

work page Pith review arXiv

[5] [7]

Im, Y.; Jo, B.; Wi, J.; Baek, S.; Min, T

Reinforcement Learning via Self- Distillation.arXiv preprint arXiv:2601.20802. Im, Y.; Jo, B.; Wi, J.; Baek, S.; Min, T. H.; Lee, J. H.; Oh, S.; Shin, I.; and Lee, S

arXiv

[6] [8]

Li, W.; Bishop, W.; Li, A.; Rawles, C.; Campbell-Ajala, F.; Tyamagundlu, D.; and Riva, O

Modular and Multi-Path- AwareOfflineBenchmarkingforMobileGUIAgents.arXiv preprint arXiv:2512.12634. Li, W.; Bishop, W.; Li, A.; Rawles, C.; Campbell-Ajala, F.; Tyamagundlu, D.; and Riva, O

arXiv

[7] [9]

Liu,Y.;Liu,Z.;Zhu,S.;Li,P.;Xie,C.;Wang,J.;Hu,X.;Han, X.; Yuan, J.; Wang, X.; et al

Ui-voyager: A self-evolving gui agent learning via failed experience.arXiv preprint arXiv:2603.24533. Liu,Y.;Liu,Z.;Zhu,S.;Li,P.;Xie,C.;Wang,J.;Hu,X.;Han, X.; Yuan, J.; Wang, X.; et al

arXiv

[8] [10]

arXiv preprint arXiv:2509.15221

Scalecua: Scaling open-source computer use agents with cross-platform data. arXiv preprint arXiv:2509.15221. Lu, Q.; Shao, W.; Liu, Z.; Du, L.; Meng, F.; Li, B.; Chen, B.;Huang,S.;Zhang,K.;andLuo,P.2025a. Guiodyssey:A comprehensive dataset for cross-app gui navigation on mo- bile devices. InProceedings of the IEEE/CVF International Conference on Computer ...

arXiv