Artificial Intelligence · arXiv:2609.32166 v1 · submitted 26 Sep 2026 · 41 references
PastForward: Faster On-Device GUI Agents via Computational Experience Reuse
The paper says
“Running GUI agents on edge devices can keep sensitive screens and interaction histories local, but the computational cost of inference at every action step makes deployment challenging.”
more of the abstract
Existing GUI agent systems either perform full vision-language model (VLM) inference at each action step or reuse coarse-grained knowledge matched to prior tasks. However, dynamic mobile environments and user tasks make it difficult to fully utilize prior task executions without additional fine-tuning or task-specific offline exploration. To address this challenge, we present PastForward, a system that accelerates GUI agents through validated, fine-grained reuse of computational experience accumulated during ordinary task execution. During decoding, PastForward retrieves prior output sequences as device-adaptive multi-token proposals and verifies them in a single VLM forward pass. Across action steps, it uses prior GUI transitions to begin next-step inference while the device executes the current action, retains the early computation only when the predicted screen matches the observed screen, and carries reusable KV states forward. We evaluate PastForward on AndroidWorld workloads derived from real mobile usage patterns using multiple VLM backbones across server and edge platforms. On device, PastForward achieves action-step latency speedups of 1.63-2.36× while maintaining task success rates.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 29 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Sound mechanism design and a credible speedup; the on-device 'maintaining SR' claim rests on single unrepeated runs and needs statistical support. This line is wrong
3 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 41 references say.
The paper says, at Section 5.2, Table 3
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The on-device success-rate parity rests on single unrepeated runs. Every cell in Table 3 (SR, prefill, decoding, action-step latency) is reported as a point value with no variance and no stated seed count, in contrast to server-side Table 2, which reports mean +/- s.d. over three random seeds.
more of objection 1
The headline claim 'while maintaining task success rates' (abstract; also claimed in Section 5.2) is supported only by these cells. Numerically, on the Galaxy S26 Ultra PastForward's SR (18.10) is below Vanilla's (19.40), and on the Jetson it is 21.70 versus 21.84. With M=348 task instances and SR around 0.19, the binomial standard error is approximately 2.1 percentage points, so a true drop of several points is indistinguishable from noise in this single run, and equally the observed small deficits do not establish parity. The paper should either report multiple seeds with confidence intervals for the on-device SRs or hedge the claim accordingly.
The paper says, at Section 3.4 (rolling-image KV carryover)
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.79):
“A retained screenshot may have originally been en- coded with an evicted screenshot and earlier action history to its left, so rolling-image KV-state carryover remains approx- imate.”
Pith says: Objection 2 of 3 · major
The 'no accuracy loss' guarantee does not cover all three mechanisms. The arguments in Section 3.2 (accepted tokens match greedy predictions) and Section 3.3 (early inference is discarded unless the observed screen matches the candidate) are correctness-preserving by construction, but the authors themselves state that rolling-image KV carryover 'remains approximate' because the RoPE re-rotation 'cannot reconstruct a changed causal context' (b195-b196).
more of objection 2
Since PastForward's prefill therefore differs from Vanilla's exactly when carryover is applied, success-rate parity cannot be recovered from mechanism guarantees for this component; it must rest on the measured SRs, which (per the Table 3 issue above) are statistically under-powered on-device. Section 5.2 should state explicitly that the approximate carryover's correctness rests on the measured results rather than on the verification argument.
The paper says, at Section 6 / Section 5.2 (transfer of the speedup)
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The authors appropriately acknowledge in Section 6 (b336-b338) that real deployments with infrequent recurrence would shrink or eliminate the benefit, so this is not a new finding. What is missing is any quantitative sensitivity of the headline 1.63-2.36x to recurrence rates below the measured ranges: the token-coverage and transition-recurrence figures in Section 2.2 (59.6-64.1% and 11.2-16.1%) are reported for the …
more of objection 3
… constructed workloads, but the paper does not show how the action-step speedup degrades as recurrence drops, for instance via a sweep over the workload-concentration parameter alpha or a synthetic recurrence-rate ablation. A short sensitivity study would substantially strengthen the transferability argument without changing the system.
4 minor objections
The paper says, at Section 5.2, Table 3 caption
Pith says: minor objection 1 of 4
The Table 3 header should state how many runs each cell averages (e.g., number of seeds and whether values are means), matching the convention used for Table 2; currently the reader cannot tell whether the point values are single runs or averaged.
The paper says, at Section 4 (implementation parameters)
Pith says: minor objection 2 of 4
The latent-index cosine-similarity threshold (0.98), the proposal-length cap (lmax=64), and the minimum match length used to report recurrence are free parameters, and only the minimum match length receives a sensitivity analysis (Figure 2a). A brief note on the sensitivity of matched-screen and proposal-acceptance rates to the 0.98 threshold would clarify how robust the reported rates are to that choice.
show 2 more
The paper says, at Abstract and Section 5.2 wording
Pith says: minor objection 3 of 4
The phrase 'while maintaining task success rates' in the abstract and Section 5.2 overstates the on-device evidence given Table 3; a hedged formulation such as 'with no statistically significant change in task success rates on the server, and comparable on-device rates' would match the data actually presented until multi-seed on-device results are available.
The paper says, at Section 5.2 (SR discussion)
Pith says: minor objection 4 of 4
The sentence at b248 ('PastForward and Vanilla achieve similar SR, with overlapping mean +/- s.d. ranges in most settings') is accurate for Table 2 and commendably careful; the on-device paragraph should apply the same evidentiary standard to Table 3 rather than asserting parity from point values.
What would settle it
Rerun the Qwen3-VL-2B on-device evaluation on the Jetson Orin Nano and Galaxy S26 Ultra (user-calibrated workload, M=348 task instances) over at least five independent seeds and report per-seed SR for PastForward and Vanilla. Compute the 90% confidence interval for the SR difference (Vanilla − PastForward) using a Wilson or bootstrap interval, and assert 'maintaining SR' only if the interval's upper bound falls within a pre-specified non-inferiority margin (e.g., ≤ 2 percentage points). If the bound exceeds the margin on either device, narrow the claim or isolate the cause with a carryover on/off ablation on the edge runtime.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 29 Sep 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
PastForward is the most convincing version of retrieval-based speculative decoding I've seen for on-device GUI agents. The core idea is genuinely new: instead of a draft model, it retrieves prior token sequences from an experience memory, verifies them in one forward pass, and accepts only what matches the current VLM's greedy predictions. That yields the standard speculative-decoding guarantee, so the adaptive decoding component provably preserves greedy output. Early inference during the A2O interval is also sound — computation is retained only after the observed screen matches the candidate. Cross-step KV carryover is approximate, but the authors say so plainly and restrict it to retained screenshots (b196). The evaluation is broad: three backbones, four platforms, AndroidWorld with three workload constructions, plus a latency breakdown that isolates each mechanism's contribution. As a system-building effort, this is well executed and honestly reported.
The soft spot is the 'maintaining task success rates' clause on-device. Table 3 reports each SR from a single run — no seeds, no variance. On the Galaxy S26 Ultra, PastForward is 18.10 vs Vanilla's 19.40; on the Jetson, 21.70 vs 21.84. With roughly 348 task instances and SR near 19%, the binomial standard error is about two points, so a genuine drop of several points cannot be distinguished from noise. The mechanism-level guarantees cover adaptive decoding and early inference, but not cross-step carryover, which the authors admit is approximate. On-device SR parity therefore rests on underpowered data. That is fixable: more runs, seed counts, and ideally an equivalence test.
The remaining concerns are milder. No code or data is shipped, which this community will rightly flag. The recurrence rates (token coverage 59.6–64.1%, transition recurrence 11.2–16.1%) come from constructed AndroidWorld workloads, not real user traces, and the authors acknowledge in §6 that infrequent recurrence would shrink or eliminate the benefit. That is a genuine transfer question, but it's stated honestly, not hidden.
This paper deserves a serious referee. The speedup claim is mechanistically sound and well measured; the SR claim needs tightening. I would send it to review with a request for seeds, variance reporting, artifact release, and an explicit statistical treatment of the SR comparison. That's an achievable revision, and the contribution is worth it.
The strongest objection
The central claim has two clauses: the 1.63-2.36x action-step speedup and 'while maintaining task success rates.' The speedup clause is well-mechanized and internally consistent across the latency breakdowns in Tables 2 and 3. The correctness clause on-device, however, rests on Table 3, where each SR cell is a single run with no variance or seed count, unlike server Table 2 which averages three seeds. On the Galaxy S26 Ultra, PastForward SR (18.10) is numerically below Vanilla (19.40); on the Jetson it is 21.70 vs 21.84. With M=348 task instances and SR near 19%, the binomial standard error is about 2 points, so a true drop of several points cannot be distinguished from noise in one unrepeated run. The 'no accuracy loss' arguments legitimately cover adaptive decoding (oracle token verification) and early inference (discard on screen mismatch), but these do not cover Cross-Step State Carryover: §3.4 admits rolling-image KV carryover remains 'approximate' because it cannot reconstruct a changed causal context (b196). PastForward's prefill therefore differs from Vanilla's, and SR parity cannot be recovered purely from mechanism guarantees; it must rest on the measured SRs. The authors acknowledge that workload-dependent recurrence limits the speedup (§6, b335-b338), but they do not acknowledge that the on-device SR claim is asserted from single runs. The unaddressed risk is a real few-point SR drop on-device, which would weaken the 'while maintaining task success rates' half of the headline result.
The reader's named weakest_assumption (transfer of the speedup to real workloads with lower recurrence) is explicitly acknowledged by the authors in §6 (b335-b338), so under the reviewing rule I do not re-flag it as a finding. Looking past that, the least secure unacknowledged link in the central claim is the on-device SR parity clause. Table 3 reports single-run SR values for all three edge devices with no standard deviation or seed count; on the Galaxy S26 Ultra the PastForward value (18.10) is numerically below Vanilla (19.40). The paper's own SR levels (about 19% at M=348) imply a binomial standard error near 2 points, so a several-point on-device SR drop cannot be ruled out. The mechanism-level 'no accuracy loss' guarantees cover adaptive decoding and early inference, but the §3.4 KV carryover is described by the authors themselves as approximate (b196), so PastForward's prefill differs from Vanilla's and SR parity cannot be established a priori; it must come from the measurements, which are single runs. This sharpens the reader's gap #2 (SR analysis) to the on-device setting, where no variance is reported at all, rather than the server setting the reader discussed — hence partial agreement. I gave credit where the design carries independent support: oracle-verified token acceptance and screen-mismatch discard are sound correctness mechanisms, and the latency accounting (prefill+decoding+A2O) is internally consistent, which is why I do not move toward REJECT. The concern is addressable and reinforces rather than overturns the CONDITIONAL verdict, so verdict_should_be is UNCHANGED.
The referee's report
If the central claims hold, halving on-device action-step latency without fine-tuning, offline exploration, or changes to the agent policy is a valuable systems contribution to privacy-preserving mobile GUI agents. The strengths are real: the single-pass verification mechanism provably preserves greedy decoding outputs with correct attribution to the speculative-decoding literature; the overhead accounting is concrete (retrieval under 5 ms, +0.1% peak RAM, mismatch-rollback cost 18-25 ms); the ablations in Figures 7-12 isolate each mechanism's contribution; and Section 6 honestly acknowledges that the speedup depends on workload recurrence. The paper's own analysis (b196, b335-b338) concedes that the rolling-image carryover is approximate and that low recurrence shrinks the benefit. What remains unresolved is the statistical grounding of the on-device success-rate parity, which currently rests on single unrepeated runs in Table 3. That is a load-bearing gap for one of the two clauses of the abstract's central claim, but it is fixable within scope by running the on-device evaluation over multiple seeds and reporting intervals.
Pith's reading of the paper (a machine's summary, not the authors')
PastForward claims that an on-device GUI agent can cut its per-action latency by 1.63-2.36x on edge hardware without fine-tuning or offline exploration, simply by reusing the fine-grained computation of past task executions. The system treats prior VLM output token sequences and prior screen-to-screen transitions as reusable 'computational experience,' retrieved and accepted only when validated by the current model or observed screen. Because each piece of reuse is checked, the agent's decisions and its task success rate are meant to be unchanged, while per-step cost roughly halves. The claimed benefit grows as experience accumulates across ordinary task runs, pointing toward agents that become faster with use rather than staying at a fixed cost.
What carries the argument. The carrying mechanism is validated fine-grained reuse indexed by local context rather than by task identity. A Decoding Experience Memory maps each final-layer decoder state to the suffix that followed it, so a retrieved multi-token proposal is accepted only for the prefix the current VLM verifies in a single forward pass, a speculative-decoding-style check whose 'draft' is the accumulated execution history instead of a separate small model. An Interaction Experience Memory stores completed transitions keyed by a perceptual screen hash plus action signature, with a latent-index fallback, so a candidate next screen can seed next-step inference early; that computation is kept only when the visual hash of the observed screen matches. Prompt-prefix and rolling-image KV-state carryover reuse prefill work between consecutive steps, rotating retained screenshot key states by the rotary position embedding rotation between old and new token positions. A runtime controller picks the proposal length per device from online estimates of token acceptance and verification latency, adapting on the fly, for example to a reduced power budget.
Core claim. On the paper's own terms, the discovery is that meaningful GUI-agent acceleration lives in recurrence below the level of tasks or trajectories: previously generated token sequences cover 59.6-64.1% of generated tokens and GUI transitions recur in 11.2-16.1% of action steps on AndroidWorld workloads, even when no exact task or trajectory repeats. PastForward converts that recurrence into latency reduction in three guarded ways: it proposes multi-token continuations retrieved from a memory of past decoding states and verifies them in a single VLM forward pass, preserving greedy decoding output; it predicts the next screen from a prior transition and starts next-step inference during the idle action-to-observation interval, discarding the work if the observed screen mismatches; and it carries KV states across steps, including position-adjusted rotary embeddings for retained screenshots. With these mechanisms the system reports action-step speedups of 1.40-1.68x on a server and 1.63-2.36x on three edge devices against a Vanilla baseline, keeping task success rates within the baseline's run-to-run noise, at about 0.1% added peak RAM. The load-bearing premise is that real mobile usage repeats tokens and screen transitions at roughly the rates measured on the benchmark workloads.
The premise it rests on. The central premise is that real phone usage repeats VLM output token sequences and screen transitions frequently enough, near the measured 59.6-64.1% token coverage and 11.2-16.1% transition recurrence on AndroidWorld workloads, that there is recurring computation to reuse in the first place.
If the paper is right (Pith's inferences, not the authors')
Each subsequent action step becomes cheaper automatically as execution records accumulate: the decoding benefit grows from a 1.037 s reduction over the first 10% of task instances to 1.606 s over the last 10%, and early-inference savings grow from 0.272 s to 0.577 s.
Because verification preserves greedy decoding and mismatched early screens are discarded, the agent's action distribution and task success rate are unchanged, so the system composes with any VLM backbone and agent policy without fine-tuning.
The gain concentrates in decoding on server-class hardware but in prefill on the Jetson and smartphones, so each device type gets a different but positive speedup, with prefill token counts cut by 70.7-77.9% for multi-screenshot backbones.
Speedups hold across runtime types (CUDA and smartphone NPU) and translate to 1.48-2.31x over FastV visual-token pruning and 1.23-1.52x over the Agent-X pipeline acceleration, while the proposal-length controller beats any fixed proposal length on the tested devices.
If the recurrence rates hold on real deployments, the natural consequence is a self-improving agent whose marginal cost per action falls with everyday use, which is exactly what makes privacy-preserving on-device agents viable where each first action step would otherwise cost seconds of local inference.
The retrieve-propose-validate-in-one-pass pattern could transfer to other latency-critical repetitive inference loops, such as code-editing assistants or form-filling agents with a stable instruction and a revisiting state space, though the paper only evaluates GUI tasks.
The dependence on the action-to-observation interval length implies the early-inference gain is largest for slow-responding apps; apps with near-instant UI feedback would hide little early compute and should show mostly the decoding gain, a split the paper does not isolate but that is directly testable.
The scheme's memory requirements appear modest (77 MiB in RAM plus screens on disk), so the reuse approach should survive on typical phone memory budgets, but eviction policies that shrink the decoding memory to 4 MiB measurably degrade the benefit, putting a floor under how small the experience store can be.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Running GUI agents on edge devices can keep sensitive screens and interaction histories local, but the computational cost of inference at every action step makes deployment challenging. Existing GUI agent systems either perform full vision-language model (VLM) inference at each action step or reuse coarse-grained knowledge matched to prior tasks. However, dynamic mobile environments and user tasks make it difficult to fully utilize prior task executions without additional fine-tuning or task-specific offline exploration. To address this challenge, we present PastForward, a system that accelerates GUI agents through validated, fine-grained reuse of computational experience accumulated during ordinary task execution. During decoding, PastForward retrieves prior output sequences as device-adaptive multi-token proposals and verifies them in a single VLM forward pass. Across action steps, it uses prior GUI transitions to begin next-step inference while the device executes the current action, retains the early computation only when the predicted screen matches the observed screen, and carries reusable KV states forward. We evaluate PastForward on AndroidWorld workloads derived from real mobile usage patterns using multiple VLM backbones across server and edge platforms. On device, PastForward achieves action-step latency speedups of 1.63-2.36× while maintaining task success rates.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-29
The authors' main claim, as the classifier picked it out 0.93
“On device, PastForward achieves action-step latency speedups of 1.63-2.36$\times$ while maintaining task success rates.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.15 · human participants 0.09 · controlled comparison 0.30 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.56 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Running GUI agents on edge devices can keep sensitive screens and interaction histories local, but the computational cost of inference at every action step makes deployment challenging. | background 1.00 | 0.06 | 0.03 | asserted 0.87 |
| Existing GUI agent systems either perform full vision-language model (VLM) inference at each action step or reuse coarse-grained knowledge matched to prior tasks. | background 0.98 | 0.04 | 0.02 | asserted 1.00 |
| However, dynamic mobile environments and user tasks make it difficult to fully utilize prior task executions without additional fine-tuning or task-specific offline exploration. | gap 0.84 | 0.05 | 0.02 | asserted 0.98 |
| To address this challenge, we present PastForward, a system that accelerates GUI agents through validated, fine-grained reuse of computational experience accumulated during ordinary task execution. | contribution 1.00 | 0.78 | 0.27 | asserted 0.99 |
| During decoding, PastForward retrieves prior output sequences as device-adaptive multi-token proposals and verifies them in a single VLM forward pass. | method 0.99 | 0.88 | 0.09 | asserted 1.00 |
| Across action steps, it uses prior GUI transitions to begin next-step inference while the device executes the current action, retains the early computation only when the predicted screen matches the observed screen, and carries reusable KV states forward. | method 1.00 | 0.84 | 0.09 | asserted 1.00 |
| We evaluate PastForward on AndroidWorld workloads derived from real mobile usage patterns using multiple VLM backbones across server and edge platforms. | method 0.85 | 0.95 | 0.05 | asserted 0.96 |
| On device, PastForward achieves action-step latency speedups of 1.63-2.36$\times$ while maintaining task success rates. | result 1.00 | 0.97 | 0.58 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-29: 1 request, 6,303 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (13)
Showing 9 of the paper's 13 figures. Figures are the authors', reproduced from the paper.
References (41)
- [11] Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2023. Fast inference from transformers via speculative decoding. InInternational conference on machine learning. PMLR, 19274–19286 on Pith
- [1] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. 2025. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631 (2025) on PithPith's reviewarXiv
- [2] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. 2024. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. InEuropean Conference on Computer Vision. Springer, 19–35 on Pith
- [3] Minghao Chen, Xinyi Hu, Zhou Yu, and Yufei Yin. 2026. AutoRPA: Efficient GUI Automation through LLM-Driven Code Synthesis from Interactions. InForty-third International Conference on Machine Learn- ing on Pith
- [4] Jinha Chung, Byeongjun Shin, Jiin Kim, and Minsoo Rhu. 2026. Agent- X: Full Pipeline Acceleration of On-device AI Agents. InProceedings of the 24th Annual International Conference on Mobile Systems, Appli- cations and Services. 144–157 on Pith
- [5] Gaole Dai, Shiqi Jiang, Ting Cao, Yuanchun Li, Yuqing Yang, Rui Tan, Mo Li, and Lili Qiu. 2026. V-Droid: Advancing Mobile GUI Agent Through Generative Verifiers. InProceedings of the 32nd Annual In- ternational Conference on Mobile Computing and Networking(Austin, TX, USA)(Mob... on Pith
- [6] Marco De Nadai, Angelo Cardoso, Antonio Lima, Bruno Lepri, and Nuria Oliver. 2019. Strategies and limitations in app usage and human mobility.Scientific reports9, 1 (2019), 10935 on Pith
- [7] Lutfi Eren Erdogan, Nicholas Lee, Siddharth Jha, Sehoon Kim, Ryan Tabrizi, Suhong Moon, Coleman Richard Charles Hooper, Gopala Anumanchipalli, Kurt Keutzer, and Amir Gholami. 2024. Tinyagent: Function calling at the edge. InProceedings of the 2024 Conference on Empirical Metho... on Pith
- [8] Runxi Huang, Liyu Zhang, Shengzhong Liu, and Xiaomin Ouyang on Pith
- [9] Jungjae Lee, Dongjae Lee, Chihun Choi, Youngmin Im, Jaeyoung Wi, Kihong Heo, Sangeun Oh, Sunjae Lee, and Insik Shin. 2025. Verisafe agent: Safeguarding mobile gui agent via logic-based action verifi- cation. InProceedings of the 31st Annual International Conference on Mobile C... on Pith
- [10] Sunjae Lee, Junyoung Choi, Jungjae Lee, Munim Hasan Wasi, Hojun Choi, Steve Ko, Sangeun Oh, and Insik Shin. 2024. Mobilegpt: Aug- menting llm with human-like app memory for mobile task automation. InProceedings of the 30th Annual International Conference on Mobile Computing an... on Pith
- [12] Kevin Qinghong Lin, Linjie Li, Difei Gao, Zhengyuan Yang, Shiwei Wu, Zechen Bai, Stan Weixian Lei, Lijuan Wang, and Mike Zheng Shou. 2025. Showui: One vision-language-action model for gui visual agent. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR... on Pith
- [13] Yozen Liu, Xiaolin Shi, Lucas Pierce, and Xiang Ren. 2019. Character- izing and forecasting user engagement with in-app action graph: A case study of snapchat. InProceedings of the 25th ACM SIGKDD inter- national conference on knowledge discovery & data mining. 2023–2031 on Pith
- [14] Xudong Lu, Yinghao Chen, Cheng Chen, Hui Tan, Boheng Chen, Yina Xie, Rui Hu, Guanxin Tan, Renshou Wu, Yan Hu, et al. 2025. Bluelm- v-3b: Algorithm and system co-design for multimodal large language models on mobile devices. In2025 IEEE/CVF Conference on Computer Vision and Pat... on Pith
- [15] Hongze Mi, Yibo Feng, WenJie Lu, Song Cao, Jinyuan Li, Yanming Li, Xuelin Zhang, Haotian Luo, Songyang Peng, He Cui, Tengfei Tian, Jun Fang, Hua Chai, and Naiqiang Tan. 2026. Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolu- tion. InForty-thi... on Pith
- [16] Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. 2025. Ui- tars: Pioneering automated gui interaction with native agents.arXiv preprint arXiv:2501.12326(2025) on PithPith's reviewarXiv
- [17] Zerui Qin, Sheng Yue, Xingyuan Hua, Yongjian Fu, and Ju Ren. 2026. Executable Agentic Memory for GUI Agent. InForty-third International Conference on Machine Learning on Pith
- [18] Qwen Team. 2026. Qwen3.5: Towards Native Multimodal Agents on Pith
- [19] Chris Rawles, Sarah Clinckemaillie, Yifan Chang, Jonathan Waltz, Gabrielle Lau, Marybeth Fair, Alice Li, William Bishop, Wei Li, Fo- lawiyo Campbell-Ajala, et al. 2025. Androidworld: A dynamic bench- marking environment for autonomous agents. InInternational Confer- ence on Le... on Pith
- [20] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing568 (2024), 127063 on Pith
show 21 more
- [21] Shizuo Tian, Hao Wen, Yuxuan Chen, Jiacheng Liu, Shanhui Zhao, Guohong Liu, Ju Ren, Yunxin Liu, and Yuanchun Li. 2026. Agentprog: Empowering long-horizon gui agents with program-guided context management. InProceedings of the 24th Annual International Confer- ence on Mobile Sy... on Pith
- [22] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2- vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191(2024) on PithPith's reviewarXiv
- [23] Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig on Pith
- [24] Hao Wen, Yuanchun Li, Guohong Liu, Shanhui Zhao, Tao Yu, Toby Jia-Jun Li, Shiqi Jiang, Yunhao Liu, Yaqin Zhang, and Yunxin Liu. 2024. Autodroid: Llm-powered task automation in android. InProceedings of the 30th annual international conference on Mobile computing and networking... on Pith
- [25] Hao Wen, Shizuo Tian, Borislav Pavlov, Wenjie Du, Yixuan Li, Ge Chang, Shanhui Zhao, Jiacheng Liu, Yunxin Liu, Ya-Qin Zhang, et al on Pith
- [26] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh J Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. 2024. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments.Advances in Neural Information Process... on Pith
- [27] Haiyang Xu, Xi Zhang, Haowei Liu, Junyang Wang, Zhaozai Zhu, Shengjie Zhou, Xuhao Hu, Feiyu Gao, Junjie Cao, Zihua Wang, et al on Pith
- [28] Autodroid-v2: Boosting slm-based gui agents via code generation. InProceedings of the 23rd Annual International Conference on Mobile Systems, Applications and Services. 223–235 on Pith
- [29] Jiabo Ye, Xi Zhang, Haiyang Xu, Haowei Liu, Junyang Wang, Zhao- qing Zhu, Ziwei Zheng, Feiyu Gao, Junjie Cao, Zhengxi Lu, et al on Pith
- [30] Christoph Zauner. 2010. Implementation and benchmarking of per- ceptual image hash functions. (2010) on Pith
- [31] Mobile-agent-v3.5: Multi-platform fundamental gui agents. arXiv preprint arXiv:2602.16855(2026) on PitharXiv
- [32] Qinglong Yang, Haoming Li, Haotian Zhao, Xiaokai Yan, Jingtao Ding, Fengli Xu, and Yong Li. 2026. FingerTip 20K: A Benchmark for Proac- tive and Personalized Mobile LLM Agents. InThe Fourteenth Interna- tional Conference on Learning Representations on Pith
- [33] Yinyuan Zhang, Daliang Xu, Zhiyang Chen, Chenghua Wang, Ying Zhang, Mengwei Xu, and Gang Huang. 2026. VLMCache: Efficient On-Device Vision-Language Model Inference. InProceedings of the 24th Annual International Conference on Mobile Systems, Applications and Services. 854–867 on Pith
- [34] Mobile-agent-v3: Fundamental agents for gui automation.arXiv preprint arXiv:2508.15144(2025) on PithPith's reviewarXiv
- [35] Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody H Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. 2024. Sglang: Efficient execution of structured lan- guage model programs.Advances in neural information processing systems37 (2... on Pith
- [36] Chi Zhang, Zhao Yang, Jiaxuan Liu, Yanda Li, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. 2025. Appagent: Multimodal agents as smartphone users. InProceedings of the 2025 CHI conference on human factors in computing systems. 1–20 on Pith
- [37] Jiwen Zhang, Ya-Qi Yu, Minghui Liao, Wentao Li, Jihao Wu, and Zhongyu Wei. 2025. UI-hawk: Unleashing the screen stream under- standing for mobile GUI agents. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 18228–18247 on Pith
- [39] Shanhui Zhao, Hao Wen, Wenjie Du, Cheng Liang, Yunxin Liu, Xi- aozhou Ye, Ye Ouyang, and Yuanchun Li. 2025. LLM-explorer: towards efficient and affordable LLM-based exploration for mobile apps. In Proceedings of the 31st Annual International Conference on Mobile Com- puting an... on Pith
- [41] Hanzhang Zhou, Xu Zhang, Panrong Tong, Jianan Zhang, Liangyu Chen, Quyu Kong, Chenglin Cai, Chen Liu, Yue Wang, Jingren Zhou, et al. 2025. Mai-ui technical report: Real-world centric foundation gui agents.arXiv preprint arXiv:2512.22047(2025). 14 on PitharXiv
- [2025] Agent Workflow Memory. InForty-second International Confer- ence on Machine Learning. 13 Park et al on Pith
- [2026] MobileExplorer: Accelerating On-Device Inference for Mobile GUI Agents via Online Exploration.arXiv preprint arXiv:2605.26546 (2026) on PithPith's reviewarXiv
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment