Pith. sign in

REVIEW 3 major objections 5 minor 78 references

A single GUI agent can match three domain specialists by distilling on the action, not the reasoning.

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

2026-08-03 09:13 UTC pith:HI7GY5QS

load-bearing objection A sensible action-aware distillation recipe whose main reported gain is statistically indistinguishable from zero, with the promised statistical tests absent from the appendix. the 3 major comments →

arxiv 2607.29320 v1 pith:HI7GY5QS submitted 2026-07-31 cs.AI

MAGA: Multi-Platform Self-Fusion of GUI Agents via Structured Action Distillation

Hang Yan , Zhangxuan GU , Beitong Zhou , Jiaxuan Chen , Runze Li , Yusong Hu , Shuheng Shen , Changhua Meng This is my paper
classification cs.AI
keywords GUI agentson-policy distillationstructured actionsmodel fusionmulti-platformvision-language modelaction-aware trainingdistillation signal allocation
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.

GUI assistants are usually trained separately for mobile, web, and desktop, forcing users and developers to juggle multiple models. The paper tries to fuse three specialist models into one student that keeps each domain's skill. It argues that the usual fusion recipes fail for a specific reason: averaging weights corrupts actions when the specialists disagree, and on-policy distillation spreads its teaching signal evenly over long reasoning traces, starving the few tokens that actually execute an action. The proposed method, MAGA, re-weights the distillation loss by the structure of the action — type versus parameters — and gives the teacher a training-only hint of the correct action type. At 8B scale, the fused model reaches 51.2% mean success rate on mobile, desktop, and web benchmarks, 2.0% above the strongest baseline, and preserves 99.9% of the teachers' average performance.

Core claim

MAGA's central claim is that the executable action is the only part of a GUI agent's response that changes the environment, so distillation should allocate signal to it according to its grammar. Each action is a type token plus type-specific parameters. MAGA drops fully correct rollouts from the loss (Rule 1), doubles the weight of the entire action span when the type is right but the parameters are wrong (Rule 2), and doubles only the type while masking the invalid parameters when the type is wrong (Rule 3). On the teacher side, the frozen domain teacher sees a hint containing only the correct action type appended to its prompt, which sharpens the log-probability scores it gives to the stud

What carries the argument

The load-bearing object is the structured action a=(z, p_z): a discrete action type z that selects a parameter schema p_z (coordinates, text, URL, keys, or empty). Three student-side token-weight rules use the binary exact-action acceptance gate to set the weight w_t for each token: fully correct actions get zero weight; correct-type/wrong-parameter actions get weight 1+β on type and parameter tokens; wrong-type actions get 1+β on the type token only, with parameter tokens masked to zero. A teacher-side hint h(z*) appends the reference action type to the domain teacher's prompt, so the teacher scores the student's sampled tokens with knowledge of the correct type. The final loss is the route

Load-bearing premise

The binary exact-action acceptance gate (coordinate distance ≤ 0.07 of the shorter screen side, token F1 > 0.5 for text, exact match for keys and empty schemas) decides which of the three training rules applies; if that rule-based evaluator labels a partially correct or semantically equivalent action as fully correct or fully wrong, Rule 1 zeroes out a trajectory that still carried useful signal, or Rules 2/3 amplify noise, and the paper does not validate the gate against hum

What would settle it

On the same 8B pipeline, replace the hand-thresholded acceptance gate with a continuous correctness score (e.g., human labels on a few thousand rollouts or a learned verifier) and re-run MAGA with the same student, teachers, and β. If mean SR does not improve over ordinary routed OPD, or if the gain over the best baseline disappears, then the reported 2.0% advantage comes from the gate's specific thresholds rather than from the action-structure re-allocation. Alternatively, keep the gate but remove the teacher hint: if TNS drops substantially below 99.9%, the hint's privileged information is t

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

If this is right

  • A single 8B model can match the average success of three domain-specialist teachers on mobile, desktop, and web tasks, which makes a one-model deployment feasible without a large accuracy penalty.
  • Weight merging is the weakest fusion strategy on high-disagreement spatial actions: on a 66-case subset, MAGA corrects 10 of the merge's 19 errors while introducing one new one, an accuracy gain of 13.6 points.
  • Action-type errors dominate (30.9% of held-out actions before training) and are the most correctable: with only the type replaced by the ground truth, the frozen student regenerates correct parameters for 68.4% of wrong-type responses.
  • The action occupies only 3.9–7.1% of response tokens, yet re-allocating distillation signal to those tokens yields consistent gains across all three domains and both model sizes.
  • Removing any single student-side rule or the teacher hint lowers at least one domain's success rate, confirming each component contributes.

Where Pith is reading between the lines

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

  • Because the three student-side rules are gated by a hand-thresholded binary evaluator, a plausible extension is to soften the gate into a confidence-weighted mixture of rules; the paper's results do not show whether the gain comes from the thresholds or from the action-aware structure itself.
  • The teacher hint is a form of privileged information (the correct type is never shown to the student). A direct test of the mechanism is to gradually corrupt or coarsen the hint — e.g., hint with a randomly chosen type — to measure how much of the 99.9% teacher-normalized score depends on exact type leakage.
  • The current training uses only single-step supervision and proprietary data; the paper itself notes failed trajectories are much longer than successful ones. A natural next experiment is to apply MAGA to multi-step rollouts with a learned or model-based acceptance gate, which would test whether action-level re-weighting also helps error recovery over long horizons.

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

3 major / 5 minor

Summary. The paper presents MAGA, a training recipe for fusing separately trained per-domain GUI agents (mobile, desktop, and web) into a single student VLM. It augments per-sample routed on-policy distillation (OPD) with two mechanisms: student-side conditional signal re-allocation, which uses the correctness of the parsed action to zero, amplify, or mask distillation weights on reasoning/action-type/parameter tokens (Section 3.2), and a teacher-side hint containing only the reference action type, which is appended to the teacher's prompt during training and never to the student's prompt (Section 3.3). The method is evaluated on Qwen3-VL-2B/8B across MobileWorld, OSWorld, and WebVoyager. The main results report the highest mean success rate among unified methods at both scales, e.g., 51.2% vs 49.2% for UI-MOPD at 8B, with TNS 99.9%, and the paper includes ablations, action-outcome analyses, teacher-matching analyses, and an oracle-type intervention study.

Significance. The action-structure insight is sound and the method is technically plausible: action tokens are the only executable interface, so focusing distillation on them makes sense, and the decomposition into type/parameter errors is natural. If the experimental claims hold, the paper provides a useful recipe for multi-platform GUI agent consolidation without weight merging. The ablations in Table 2 are internally consistent and support the contribution of the student-side rules and the teacher-side hint; the oracle-type intervention (Section 4.7) is a clever diagnostic showing that many wrong actions are recoverable from type correction alone. However, the central superiority claim is currently under-supported: Table 1 has no confidence intervals or significance tests, despite Section 4.1's explicit promise of Appendix B tests, and the implied task counts make the 2.0-point mean gain over UI-MOPD statistically indistinguishable from zero. The binary exact-action gate (Table 6) is another unvalidated decision point. These are fixable concerns, but they are load-bearing for the abstract's headline claims. The proprietary training data and unreleased artifacts also limit reproducibility.

major comments (3)
  1. [§4.1, Appendix B, Table 1] The paper states in §4.1 that 'statistical tests are provided in Appendix B,' but Appendix B contains no tests, confidence intervals, or p-values; it only restates metric definitions. Table 1's 8B results are point estimates from 117+369+140 tasks. The implied success counts for MAGA vs UI-MOPD (40 vs 36, 167 vs 165, 104 vs 101) give per-domain differences of 3.4, 0.6, and 2.2 points, with approximate SEs of 6.1, 3.7, and 5.3 points; the mean 2.0-point advantage has an approximate SE of 3.0. The headline 'outperforms the strongest baseline' is therefore unsupported by the reported evidence. Please add task-level tests (McNemar/permutation) and bootstrap CIs for both scales and all domains, or temper the claim. The same gap appears in §4.6/Appendix A.5, where bootstrap resampling is mentioned but no intervals are reported.
  2. [§3.2, §A.4, Table 6] The binary exact-action acceptance gate is load-bearing: Rule 1 zeros the entire distillation signal when the gate accepts, while Rules 2/3 select the token weighting when it rejects. Table 6's correctness rules use hand-set thresholds (0.07 normalized distance, token F1>0.5, etc.) and produce a binary decision with no partial credit. The paper does not validate this gate against human judgments, task-level success, or alternative thresholds. A partially correct but semantically valid action that fails the gate is treated as fully erroneous, which can amplify correction pressure on already-correct parts; a spurious acceptance suppresses useful signal. Please add a validation or sensitivity analysis for the gate and report rule firing frequencies on training data.
  3. [§A.1, §4.9] Reproducibility is limited: the 343k-example training set is proprietary and unreleased, no code link or checkpoints are provided, and only the WebVoyager subset IDs are listed. Since the central contribution is an empirical training recipe, the authors should release as much as possible (parsing/evaluator code, thresholds, bootstrap scripts, and, ideally, a public-data reimplementation) so that the 2.0% claim can be verified or refuted externally.
minor comments (5)
  1. [Table 1] Please report the raw success counts (e.g., 40/117) underlying the percentages; rounding makes exact counts ambiguous and prevents readers from computing their own confidence intervals.
  2. [Figure 1(a) and §4.4] Figure 1(a) uses '900 random samples' and §4.4 uses '900 held-out test examples'; clarify whether these are the same 900 tasks.
  3. [References] SEED and OPID each appear twice in the bibliography (Wu et al. 2026a/2026b; Yang et al. 2026b/2026c); merge the duplicates.
  4. [Abstract/Contributions] The spelling alternates between 'Maga' and 'MAGA'; use a single spelling throughout.
  5. [Appendix D] The URL parameter class contains only 5 examples; the paper correctly declines to draw conclusions there, but this limitation should be mentioned when param-class results are cited.

Circularity Check

0 steps flagged

No significant circularity: empirical claims rest on external benchmarks; the Appendix B statistics gap is missing evidence, not circularity.

full rationale

The paper's central claims are empirical: MAGA is a training recipe whose headline numbers (51.2 mean SR at 8B, TNS 99.9%) are measured on fixed external benchmarks (MobileWorld 117 tasks, OSWorld 369, WebVoyager 140). No equation in the paper constructs the reported SR from the training signal; the student is evaluated by task-level success, not by the exact-action gate used during training. The training-time use of the reference action in the student-side allocation and teacher-side hint is an explicit supervised label, and the paper states the student never receives the hint (PS(x)=P(x)), so there is no fitted parameter renamed as a prediction. Disabling both sides reduces the objective to ordinary routed OPD (Eq. 4 reduces to Eq. 2), making the comparison self-contained. Self-citations appear in related work (UI-Venus, VenusBench, GUI-G2, etc.) but none is load-bearing for the method's derivation. The one flagged defect is evidentiary, not circular: Section 4.1 says 'statistical tests are provided in Appendix B,' but Appendix B contains no tests, confidence intervals, or p-values; that is a missing-support/correctness-risk issue, not a reduction of the result to its inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The core innovation is a weighting scheme. Its free parameters are β and the acceptance thresholds; neither is derived from theory. The main assumptions are that actions are structured as type+params, teachers are reliable, reference actions are known, and the rule-based gate is an adequate correctness oracle.

free parameters (2)
  • β (amplification coefficient) = 1.0
    Chosen by hand and fixed across all experiments; controls how much action/type tokens are upweighted relative to reasoning tokens. Not tuned per domain or scale.
  • Action acceptance thresholds = 0.07 / 0.14 / token F1 > 0.5
    Table 6's rules define what counts as a fully correct action and therefore determine which distillation rule applies; thresholds are chosen by hand without justification.
axioms (4)
  • domain assumption GUI actions have a structured representation a=(z, p_z) with a discrete action type and type-specific parameters shared across mobile, web, and desktop domains.
    Introduced in §3.1 and used by all three student-side rules; if domains use incompatible action grammars, the method does not apply.
  • domain assumption Frozen domain-specific teachers provide reliable per-token log-probability supervision for their own domain.
    Routed OPD and the teacher-side hint rely on the teacher's log-probabilities being meaningful training signals; weak or miscalibrated teachers would corrupt the advantage.
  • domain assumption Ground-truth reference actions a* are available for all training samples and are correct.
    The student-side rules and the teacher hint both depend on knowing the correct action type and parameters; incorrect labels would mis-allocate distillation signal.
  • ad hoc to paper The exact-action acceptance gate accurately identifies fully correct actions.
    Table 6 and §A.4 define the gate; Rule 1 zeroes out the entire trajectory when the gate says the action is fully correct. If the gate is noisy, the method suppresses useful signal or amplifies noise.

pith-pipeline@v1.3.0-daily-deepseek · 16882 in / 12456 out tokens · 130659 ms · 2026-08-03T09:13:29.764502+00:00 · methodology

0 comments
read the original abstract

Graphical user interface (GUI) agents based on large language models are increasingly deployed across mobile, web, and desktop environments. However, existing agents are typically domain-specific, limiting the deployment and user experience. This motivates the consolidation of specialized models into a single cross-environment policy. Weight merging directly merges domain-specific experts but can corrupt executable actions under expert disagreement, while on-policy distillation (OPD) avoids conflicting teacher supervision yet still treats all response tokens equally during distillation, ignoring that action tokens are the only interface between the environment and the agent. To address this, We introduce MAGA that re-allocates training signal according to the structured action. Based on the correctness of the generated action, it suppresses unnecessary or invalid distillation signals and focuses learning on erroneous actions. Besides, a training-only hint optimizes the supervision signal provided by domain-specific teachers without changing the student input. Across two model scales, MAGA achieves the highest mean success rate, outperforming the strongest baseline by 2.0% at 8B and achieves almost the same average performance with teachers.

Figures

Figures reproduced from arXiv: 2607.29320 by Beitong Zhou, Changhua Meng, Hang Yan, Jiaxuan Chen, Runze Li, Shuheng Shen, Yusong Hu, Zhangxuan Gu.

Figure 1
Figure 1. Figure 1: Results from three benchmarks, including Mobile [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of three strategies for unifying domain-specific GUI agents. (a) Parameter merging combines the weights [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Based on the trained student, we use 900 random samples (300 per domain) to generate response. (a) Each action [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: For initially incorrect actions, correcting solely the [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

78 extracted references · 15 linked inside Pith

  1. [2]

    Learn where to Click from Yourself: On-Policy Self-Distillation for GUI Grounding , author =

  2. [3]

    Task Arithmetic (Editing Models with Task Vectors) , author =

  3. [4]

    TIES-Merging: Resolving Interference When Merging Models , author =

  4. [5]

    Branch-Train-Merge , author =

  5. [6]

    Branch-Train-Mix (BTX) , author =

  6. [7]

    MiniLLM: Knowledge Distillation of Large Language Models (reverse-KL) , author =

  7. [8]

    DAgger: A Reduction of Imitation Learning to No-Regret Online Learning , author =

  8. [9]

    GUI Agents (grounding / mobile navigation / web / OS control) , author =

  9. [12]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Os-genesis: Automating gui agent trajectory construction via reverse task synthesis , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

  10. [13]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Seeclick: Harnessing gui grounding for advanced visual gui agents , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

  11. [14]

    International Conference on Learning Representations , volume =

    OS-ATLAS: Foundation action model for generalist GUI agents , author =. International Conference on Learning Representations , volume =

  12. [15]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages =

    Agentstore: Scalable integration of heterogeneous agents as specialized generalist computer assistant , author =. Findings of the Association for Computational Linguistics: ACL 2025 , pages =

    2025

  13. [20]

    Advances in Neural Information Processing Systems , volume =

    Webshop: Towards scalable real-world web interaction with grounded language agents , author =. Advances in Neural Information Processing Systems , volume =

  14. [27]

    International conference on machine learning , pages =

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time , author =. International conference on machine learning , pages =. 2022 , organization =

    2022

  15. [28]

    Advances in neural information processing systems , volume =

    Ties-merging: Resolving interference when merging models , author =. Advances in neural information processing systems , volume =

  16. [29]

    arXiv preprint arXiv:2606.14971 , year =

    FastMix: Fast Data Mixture Optimization via Gradient Descent , author =. arXiv preprint arXiv:2606.14971 , year =

    arXiv

  17. [30]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages =

    DaMo: Data Mixing Optimizer in Fine-tuning Multimodal LLMs for Mobile Phone Agents , author =. Findings of the Association for Computational Linguistics: ACL 2026 , pages =

    2026

  18. [31]

    arXiv preprint arXiv:2407.21783 , year =

    The llama 3 herd of models , author =. arXiv preprint arXiv:2407.21783 , year =

    Pith/arXiv arXiv

  19. [33]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Mobileworld: Benchmarking autonomous mobile agents in agent-user interactive and mcp-augmented environments , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

  20. [34]

    Advances in Neural Information Processing Systems , volume =

    Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments , author =. Advances in Neural Information Processing Systems , volume =

  21. [35]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Webvoyager: Building an end-to-end web agent with large multimodal models , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

  22. [36]

    Thinking Machines Lab: Connectionism , year =

    Kevin Lu and Thinking Machines Lab , title =. Thinking Machines Lab: Connectionism , year =

  23. [40]

    International Conference on Learning Representations , volume =

    Minigpt-4: Enhancing vision-language understanding with advanced large language models , author =. International Conference on Learning Representations , volume =

  24. [41]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages =

    A survey of state of the art large vision language models: Benchmark evaluations and challenges , author =. Proceedings of the Computer Vision and Pattern Recognition Conference , pages =

  25. [42]

    Advances in Neural Information Processing Systems , volume =

    What matters when building vision-language models? , author =. Advances in Neural Information Processing Systems , volume =

  26. [43]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages =

    Efficient inference for large vision-language models: Bottlenecks, techniques, and prospects , author =. Findings of the Association for Computational Linguistics: ACL 2026 , pages =

    2026

  27. [44]

    Computation , volume =

    A comprehensive survey and guide to multimodal large language models in vision--language tasks , author =. Computation , volume =. 2026 , publisher =

    2026

  28. [47]

    Forty-third International Conference on Machine Learning , year =

    VenusBench-Mobile: A Challenging and User-Centric Benchmark for Mobile GUI Agents with Capability Diagnostics , author =. Forty-third International Conference on Machine Learning , year =

  29. [48]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    GUI-G ^2 : Gaussian Reward Modeling for GUI Grounding , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =

  30. [50]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    GUI-SAGE: Enhancing GUI Automation with Self-Explanatory Learning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  31. [51]

    2026 , eprint =

    OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning , author =. 2026 , eprint =

    2026

  32. [52]

    2026 , eprint =

    SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning , author =. 2026 , eprint =

    2026

  33. [55]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Beyond Layer-Wise Merging: Chain-of-Merging for Vision-Language Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  34. [56]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year =

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author =. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year =

  35. [57]

    Advances in Neural Information Processing Systems , volume =

    Cofft: Chain of foresight-focus thought for visual language models , author =. Advances in Neural Information Processing Systems , volume =

  36. [58]

    Cheng, K.; Li, Z.; Ma, Z.; Chen, N.; Cao, J.; Sun, Q.; Ding, Z.; Xu, F.; Yan, H.; Chen, J.; et al. 2026. OpenMobile: Building open mobile agents with task and trajectory synthesis. arXiv preprint arXiv:2604.15093

    Pith/arXiv arXiv 2026

  37. [59]

    Cheng, K.; Sun, Q.; Chu, Y.; Xu, F.; YanTao, L.; Zhang, J.; and Wu, Z. 2024. Seeclick: Harnessing gui grounding for advanced visual gui agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 9313--9332

    2024

  38. [60]

    Gong, Y.; Cai, Z.; Dai, S.; Zhou, Y.; Gu, Z.; Meng, C.; and Shen, S. 2026. VenusBench-Mobile: A Challenging and User-Centric Benchmark for Mobile GUI Agents with Capability Diagnostics. In Forty-third International Conference on Machine Learning

    2026

  39. [61]

    Gu, Z.; Zeng, Z.; Xu, Z.; Zhou, X.; Shen, S.; Liu, Y.; Zhou, B.; Meng, C.; Xia, T.; Chen, W.; et al. 2025. Ui-venus technical report: Building high-performance ui agents with rft. arXiv preprint arXiv:2508.10833

    Pith/arXiv arXiv 2025

  40. [62]

    He, H.; Yao, W.; Ma, K.; Yu, W.; Dai, Y.; Zhang, H.; Lan, Z.; and Yu, D. 2024. Webvoyager: Building an end-to-end web agent with large multimodal models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 6864--6890

    2024

  41. [63]

    Hu, J.; Liu, J.; Lai, J.; Hu, J.; Sheng, Y.; Chen, S.; Li, J.; Du, D.; and Guo, S. 2026. GUI Agents with Reinforcement Learning: Toward Digital Inhabitants. arXiv preprint arXiv:2604.27955

    Pith/arXiv arXiv 2026

  42. [64]

    Jia, C.; Luo, M.; Dang, Z.; Sun, Q.; Xu, F.; Hu, J.; Xie, T.; and Wu, Z. 2025. Agentstore: Scalable integration of heterogeneous agents as specialized generalist computer assistant. In Findings of the Association for Computational Linguistics: ACL 2025, 8908--8934

    2025

  43. [65]

    Kong, Q.; Zhang, X.; Yang, Z.; Gao, N.; Liu, C.; Tong, P.; Cai, C.; Zhou, H.; Zhang, J.; Chen, L.; et al. 2026. Mobileworld: Benchmarking autonomous mobile agents in agent-user interactive and mcp-augmented environments. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 6142--6167

    2026

  44. [66]

    H.; Gonzalez, J

    Kwon, W.; Li, Z.; Zhuang, S.; Sheng, Y.; Zheng, L.; Yu, C. H.; Gonzalez, J. E.; Zhang, H.; and Stoica, I. 2023. Efficient Memory Management for Large Language Model Serving with PagedAttention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles

    2023

  45. [67]

    Li, J.; Deng, K.; Wang, Y.; Huang, J.; Shi, Y.; Tan, Q.; Lu, J.; and Liu, N. 2026 a . Online Skill Learning for Web Agents via State-Grounded Dynamic Retrieval. arXiv preprint arXiv:2606.04391

    Pith/arXiv arXiv 2026

  46. [68]

    Li, Y.; Zuo, Y.; He, B.; Zhang, J.; Xiao, C.; Qian, C.; Yu, T.; Gao, H.-a.; Yang, W.; Liu, Z.; et al. 2026 b . Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe. arXiv preprint arXiv:2604.13016

    Pith/arXiv arXiv 2026

  47. [69]

    Li, Z.; Wu, X.; Du, H.; Liu, F.; Nghiem, H.; and Shi, G. 2025. A survey of state of the art large vision language models: Benchmark evaluations and challenges. In Proceedings of the Computer Vision and Pattern Recognition Conference, 1587--1606

    2025

  48. [70]

    Lian, N.; Chen, A.; Yu, Z.; Duan, C.; Liu, F.; Liu, H.; Fu, P.; Luan, J.; Wang, Y.; Xia, S.-T.; and Wang, J. 2026. UI-MOPD : Multi-Platform On-Policy Distillation for Continual GUI Agent Learning. arXiv:2607.04425

    Pith/arXiv arXiv 2026

  49. [71]

    X.; Tian, P.; Yin, C

    Liang, C. X.; Tian, P.; Yin, C. H.; Yua, Y.; Wei, A.-H.; Li, M.; Song, X.; Wang, T.; Bi, Z.; Liu, M.; et al. 2026. A comprehensive survey and guide to multimodal large language models in vision--language tasks. Computation, 14(6): 125

    2026

  50. [72]

    Liu, Z.; Xie, J.; Ding, Z.; Li, Z.; Yang, B.; Wu, Z.; Wang, X.; Sun, Q.; Liu, S.; Wang, W.; et al. 2025. Scalecua: Scaling open-source computer use agents with cross-platform data. arXiv preprint arXiv:2509.15221

    arXiv 2025

  51. [73]

    Lu, K.; and Lab, T. M. 2025. On-Policy Distillation. Thinking Machines Lab: Connectionism. Https://thinkingmachines.ai/blog/on-policy-distillation

    2025

  52. [74]

    Q.; Zhang, X.; Jie, Z.; Sun, P.; Jin, X.; and Li, H

    Luong, T. Q.; Zhang, X.; Jie, Z.; Sun, P.; Jin, X.; and Li, H. 2024. Reft: Reasoning with reinforced fine-tuning. arXiv preprint arXiv:2401.08967

    Pith/arXiv arXiv 2024

  53. [75]

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv. org/abs/2402.03300, 2(3): 5

    Pith/arXiv arXiv 2024

  54. [76]

    Sun, Q.; Cheng, K.; Ding, Z.; Jin, C.; Wang, Y.; Xu, F.; Wu, Z.; Jia, C.; Chen, L.; Liu, Z.; et al. 2025. Os-genesis: Automating gui agent trajectory construction via reverse task synthesis. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 5555--5579

    2025

  55. [77]

    Tang, F.; Gu, Z.; Lu, Z.; Liu, X.; Shen, S.; Meng, C.; Wang, W.; Zhang, W.; Shen, Y.; Lu, W.; et al. 2026 a . GUI-G ^2 : Gaussian Reward Modeling for GUI Grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 33214--33222

    2026

  56. [78]

    Tang, F.; Gu, Z.; Lu, Z.; Zhang, S.; Zeng, Z.; Shen, S.; Meng, C.; Yan, Y.; Zhang, W.; Shen, Y.; et al. 2026 b . GUI-SAGE: Enhancing GUI Automation with Self-Explanatory Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 13007--13016

    2026

  57. [79]

    Team, V.; Gao, C.; Gu, Z.; Liu, Y.; Qiu, X.; Shen, S.; Wen, Y.; Xia, T.; Xu, Z.; Zeng, Z.; et al. 2026. Ui-venus-1.5 technical report. arXiv preprint arXiv:2602.09082

    arXiv 2026

  58. [80]

    Y.; Roelofs, R.; Gontijo-Lopes, R.; Morcos, A

    Wortsman, M.; Ilharco, G.; Gadre, S. Y.; Roelofs, R.; Gontijo-Lopes, R.; Morcos, A. S.; Namkoong, H.; Farhadi, A.; Carmon, Y.; Kornblith, S.; et al. 2022. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, 23965--23998. PMLR

    2022

  59. [82]

    Wu, J.; Yang, S.; Lu, Z.; Zhang, F.; Shen, Y.; Feng, L.; Luo, H.; Lian, Z.; Zhang, S.; Wen, Z.; et al. 2026 b . SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning. arXiv preprint arXiv:2607.14777

    Pith/arXiv arXiv 2026

  60. [83]

    P.; et al

    Wu, Z.; Wu, Z.; Xu, F.; Wang, Y.; Sun, Q.; Jia, C.; Cheng, K.; Ding, Z.; Chen, L.; Liang, P. P.; et al. 2025. OS-ATLAS: Foundation action model for generalist GUI agents. In International Conference on Learning Representations, volume 2025, 5090--5108

    2025

  61. [84]

    Xiao, B.; Xia, B.; Yang, B.; Gao, B.; Shen, B.; Zhang, C.; He, C.; Lou, C.; Luo, F.; Wang, G.; et al. 2026. Mimo-v2-flash technical report. arXiv preprint arXiv:2601.02780

    Pith/arXiv arXiv 2026

  62. [85]

    J.; Cheng, Z.; Shin, D.; Lei, F.; et al

    Xie, T.; Zhang, D.; Chen, J.; Li, X.; Zhao, S.; Cao, R.; Hua, T. J.; Cheng, Z.; Shin, D.; Lei, F.; et al. 2024. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing Systems, 37: 52040--52094

    2024

  63. [86]

    Xu, A.; Lin, B.; Xue, B.; Wang, B.; Xu, B.; Wu, B.; Zhang, B.; Lin, C.; Dong, C.; Ling, C.; et al. 2026 a . Deepseek-v4: Towards highly efficient million-token context intelligence. arXiv preprint arXiv:2606.19348

    arXiv 2026

  64. [87]

    Xu, H.; Zhang, X.; Liu, H.; Wang, J.; Zhu, Z.; Zhou, S.; Hu, X.; Gao, F.; Cao, J.; Wang, Z.; et al. 2026 b . Mobile-agent-v3. 5: Multi-platform fundamental gui agents. arXiv preprint arXiv:2602.16855

    arXiv 2026

  65. [88]

    Xue, T.; Peng, C.; Huang, M.; Guo, L.; Han, T.; Wang, H.; Wang, J.; Zhang, X.; Yang, X.; Zhao, D.; et al. 2026. Evocua: Evolving computer use agents via learning from scalable synthetic experience. arXiv preprint arXiv:2601.15876

    arXiv 2026

  66. [89]

    A.; and Bansal, M

    Yadav, P.; Tam, D.; Choshen, L.; Raffel, C. A.; and Bansal, M. 2023. Ties-merging: Resolving interference when merging models. Advances in neural information processing systems, 36: 7093--7115

    2023

  67. [90]

    Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    Pith/arXiv arXiv 2025

  68. [91]

    Yang, B.; Jin, K.; Wu, Z.; Liu, Z.; Sun, Q.; Li, Z.; Xie, J.; Liu, Z.; Xu, F.; Cheng, K.; et al. 2026 a . Os-symphony: A holistic framework for robust and generalist computer-using agent. arXiv preprint arXiv:2601.07779

    arXiv 2026

  69. [93]

    Yang, S.; Wu, J.; Lu, Z.; Shen, Y.; Zhang, F.; Feng, L.; Zhang, S.; Luo, H.; Lian, Z.; Wen, Z.; et al. 2026 c . OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning. arXiv preprint arXiv:2606.26790

    Pith/arXiv arXiv 2026

  70. [94]

    Yao, S.; Chen, H.; Yang, J.; and Narasimhan, K. 2022. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35: 20744--20757

    2022

  71. [95]

    Zhang, J.; Ji, Y.; Ren, F.; Li, Y.; Zeng, B.; Chen, Z.; Chen, K.; Shou, L.; Chen, G.; and Li, H. 2026 a . Efficient inference for large vision-language models: Bottlenecks, techniques, and prospects. In Findings of the Association for Computational Linguistics: ACL 2026, 21036--21066

    2026

  72. [96]

    Zhang, X.; Dong, Y.; Zhang, L.; Jia, C.; Dang, Z.; Fernando, B.; Liu, J.; and Shou, M. Z. 2026 b . Cofft: Chain of foresight-focus thought for visual language models. Advances in Neural Information Processing Systems, 38: 106182--106201

    2026

  73. [97]

    Zhang, X.; Dong, Y.; Zhang, L.; Jia, C.; Dang, Z.; Yao, Y.; Wu, Y.; Fernando, B.; and Liu, J. 2026 c . Beyond Layer-Wise Merging: Chain-of-Merging for Vision-Language Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 24279--24289

    2026

  74. [98]

    Zhang, Y.; Tang, S.; Li, Z.; Han, Z.; and Tresp, V. 2026 d . WebArbiter: A Principle-Guided Reasoning Process Reward Model for Web Agents. arXiv preprint arXiv:2601.21872

    Pith/arXiv arXiv 2026

  75. [99]

    Zheng, B.; Gou, B.; Kil, J.; Sun, H.; and Su, Y. 2024. Gpt-4v (ision) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614

    Pith/arXiv arXiv 2024

  76. [100]

    Zhou, B.; Huang, Z.; Guo, Y.; Gu, Z.; Xia, T.; Luo, Z.; Tang, F.; Kong, D.; Shang, Y.; Ou, S.; et al. 2025 a . VenusBench-GD: A Comprehensive Multi-Platform GUI Benchmark for Diverse Grounding Tasks. arXiv preprint arXiv:2512.16501

    arXiv 2025

  77. [101]

    Zhou, H.; Zhang, X.; Tong, P.; Zhang, J.; Chen, L.; Kong, Q.; Cai, C.; Liu, C.; Wang, Y.; Zhou, J.; et al. 2025 b . MAI-UI Technical Report: Real-World Centric Foundation GUI Agents. arXiv preprint arXiv:2512.22047

    arXiv 2025

  78. [102]

    Zhu, D.; Shen, X.; Li, X.; Elhoseiny, M.; et al. 2024. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In International Conference on Learning Representations, volume 2024, 18378--18394

    2024