Pith.

Distributed, Parallel, and Cluster Computing · arXiv:2609.32130 v1 · submitted 26 Sep 2026 · 48 references

REBASE: Device-Cloud Experience Coherence for GUI Agents Across App Updates

Beining Wu, Jun Huang, Yanxiao Zhao

The paper says

“A graphical user interface (GUI) agent that ships on a phone runs a small model and reuses experience: action paths recorded on earlier runs and cached from the cloud. When an app updates, part of this experience becomes silently wrong.”

more of the abstract

On nine real app version pairs, an agent carrying experience recorded on the old version succeeds 10.1 points less often than one carrying none, and it does not notice the mismatch until 3 steps after acting on it. We propose REBASE, a protocol that keeps the device copy and the cloud copy of the experience coherent across updates. When a version changes, the cloud replays its copy on the new version and the device verifies each recorded step before executing it; when a step still fails, the device sends the cloud evidence in increasing size, starting from an accessibility-subtree diff, until the cloud re-derives an executable patch, keyed by version so that one repair serves the whole fleet. On two Jetson devices, REBASE restores the success rate of stale experience to that of fresh experience recorded on the new version within one episode, at 708x fewer bytes and 47.3x fewer cloud calls than uploading screenshots to the cloud, and 2.5x less device energy than running without experience.

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. A well-measured systems paper on a real deployment failure mode: the central claims hold on its testbed, but the 708x byte-savings headline is conditional on an accessibility-diff geometry that is never ablated or held out. This line is wrong

2 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 48 references say.

The paper says, at Section IV; Section II-C; Section III-C

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

The version-pair selection rule in Section IV—v' is chosen as the next F-Droid release that changes an accessibility tree along a recorded path—conditions Obs. 2's 79.1% repair rate and the ladder's byte savings on the update breakage being reachable through the fixed diff geometry (root at the grandparent, three levels, one sibling hop, Section III-C).

more of objection 1

The paper does not report a per-app distribution of diff repair rates, does not ablate the diff geometry, and does not hold out app pairs selected without the accessibility-tree-change criterion. Fig. 1 already shows the crop is the knee for two of six cloud-visible apps, so the diff is not uniformly sufficient even inside the curated set. Because the 708x byte reduction flows directly from closing most events at the fingerprint or diff level, the claim that REBASE achieves that reduction for GUI agents across app updates in general is not yet established for updates that break elements outside the fixed window. Please add per-app repair-rate tables, an ablation varying root depth and sibling hops, and at least a small held-out set of version pairs selected without the tree-change criterion.

The paper says, at Section III-C, Eq. (4)

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The expected-byte model in Eq. (4) is evaluated with conditional success probabilities q_l = Pr[R(xi_l) != bottom | lower levels failed] that are estimated from the same nine version pairs used for the headline evaluation. If the 708x byte savings or the statement that the ladder 'closes most events at xi_0 or xi_1' is justified through Eq.

more of objection 2

(4), then part of the cost claim is model-fit rather than independently measured. Table I reports measured uplink and downlink kilobytes, which is the right basis, but the derivation in Section III-C mixes those measured totals with a fitted expectation. Please report the per-level measured resolution rates and measured per-event byte totals separately from the Eq. (4) prediction, with standard errors for q_l, so that the reader can see how much of the byte claim is validated by direct measurement.

5 minor objections

The paper says, at Abstract and Section V-B

Pith says: minor objection 1 of 5

The abstract's 'within one episode' is stronger than the reported data: Section V-B states first-episode recovery for 91.1% of cloud-visible units and 68.9% of blind units. Please qualify the abstract and conclusion with these rates or report the full T_r distribution instead of a median.

The paper says, at Table I

Pith says: minor objection 2 of 5

The ablation block lists X1, X2, X3, X4, X6, X7, X8 but skips X5. Please either renumber the rows or explain in the text why X5 is absent; as printed, the gap looks like an omission.

show 3 more

The paper says, at Section II-C and Fig. 1

Pith says: minor objection 3 of 5

Obs. 2 and Fig. 1 (right) are described as covering failing steps, but the caption restricts the five-evidence comparison to six cloud-visible apps. Please clarify whether and how the three blind pairs contribute to the 79.1% repair rate, or state explicitly that Obs. 2 is computed only on the six cloud-visible pairs.

The paper says, at Section IV

Pith says: minor objection 4 of 5

For reproducibility, please list the nine app version pairs, the five tasks per pair, and the selection dates from F-Droid Archive. The current text gives only the selection rule, which is necessary but not sufficient to reconstruct the testbed.

The paper says, at Section III-E and Conclusion

Pith says: minor objection 5 of 5

The paper's final paragraph notes that every measured failure exposed an accessibility tree and that screens without one start at a crop. I recommend adding a short limitation paragraph stating explicitly that the reported byte savings apply to the accessibility-tree-present regime and that the crop fallback's frequency in real deployments is unmeasured.

What would settle it

Take a held-out set of 20 real app version pairs not used in designing the diff geometry (e.g., a random sample from F-Droid release history that excludes the nine tested apps), run the REBASE ladder with the exact same diff parameters (grandparent root, three levels, one sibling hop) without re-tuning, and measure (i) the share of failing steps closed at level L1 (diff) versus L2/L3, and (ii) the resulting uplink bytes per episode and Succ@K over K=5 episodes with three seeds. If the L1 close rate falls below about 50%, or uplink bytes per episode rise by more than 5x relative to the reported 2.5 KB median, the 708x byte-savings claim is not representative of real update breakage and the central coherence claim weakens.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 29 Sep 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 37 s · 14,257 tokens read, 4,090 written
  3. Editor wrote the verdict above 36 s · 15,271 tokens read, 4,530 written
  4. Referee major revision: 2 major objections, 5 minor 57 s · 14,854 tokens read, 5,944 written

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

Short version: REBASE is a genuine systems contribution, and on its reported testbed the central claims hold. The 708x byte reduction is real but conditional on an accessibility-subtree diff geometry that is a tuned free parameter on nine curated app pairs, never ablated and never held out. That is a generalization question, not an error in what was measured.

What's new: prior work repairs locators after test-script failure, offloads screenshots on failure, filters trajectories in the cloud, or ships edits down offline. None treats an app update as a coherence event between a device copy and a cloud copy of the same experience. REBASE's assembly—update clock (cloud replays on the new version, device verifies each step) plus failure clock (escalating evidence: fingerprint, a11y diff, crop, screenshot until the cloud re-derives a version-keyed patch)—is new and is measured end-to-end on real hardware.

Credit where earned: two Jetson AGX Orin boards plus a Thor gateway running AndroidWorld; eight ablations that make each component's contribution legible; energy on real INA3221 rails; and a genuine honesty check in the three blind pairs where the cloud cannot get the new version. The headline statistic is clean: Succ@K gains 56.1 points over stale experience (CI [51.6, 60.6]) and matches fresh experience (difference 0.4, CI [-2.1, 3.0], includes zero) (b125).

The main soft spot, and the stress-test note gets it right: the diff geometry (grandparent root, three levels, one sibling hop) is specified but never ablated, and the nine pairs were selected because the update changes an accessibility tree along a recorded path (Section IV). That selection guarantees the diff has something to latch onto; it does not measure how often real updates break steps outside that geometry. The paper's own knee result shows heterogeneity even in the curated set—two of six cloud-visible apps close at the crop, not the diff (b49). So the 708x figure is conditional on a geometry that may be overfit. This bounds external validity rather than invalidating the measured result, but the paper never flags it because it has no limitations section; the selection rule appears only in the methods. The crop fallback for treeless screens is covered (Section III-E), but real-update breakage outside the diff-repairable class is unmeasured.

Smaller items: no code or data shipped; three seeds and nine app pairs; the 80x fleet dedup is trace-driven model extrapolation, not a measured fleet outcome; the conditional probabilities ql in Eq (4) are empirical rates from the same evaluation set, a mild circularity in the expected-bytes figure. And a large share of the references are the authors' own prior work—not a flaw per se, but noticeable.

Recommendation: send it to peer review. It is a solid systems paper for the deployed-GUI-agent community, the concerns are addressable conditions (release artifacts, enlarge the app and seed set, ablate the diff geometry, sensitivity analysis on the selection rule, state the bound explicitly), and a revision that settles the geometry question would make it definitive. Worth referee time.

The strongest objection

The central cost-and-recovery claims rest on the accessibility-subtree diff being the right repair signal: Obs. 2 states that a diff repairs 79.1% of failing steps (b47), and the ladder's byte advantage over screenshots (708x fewer) flows from closing most events at the fingerprint or diff level (b94). But the diff geometry is specified only as rooted at the expected element's grandparent, three levels deep with one sibling hop (Section III-C), and the nine version pairs were deliberately selected because their update changes an accessibility tree along a recorded path (Section IV). That selection guarantees the diff has something to latch onto; it does not measure the rate at which arbitrary real updates produce breakage reachable through that exact geometry. The paper's own knee result (b49) shows the diff is not even uniformly sufficient within the curated set: for two of six cloud-visible apps the knee is the crop, meaning the diff alone adds little for those apps and the ladder escalates to higher evidence. The paper reports no per-app repair-rate distribution, no ablation that fixes the diff geometry at a different tree depth, and no held-out app pairs. Consequently, the headline 708x byte reduction and the claim of matching fresh experience within one episode are conditional on a geometry that may be overfitted to the testbed; an update that shifts elements outside the grandparent-rooted three-level window would degrade the ladder toward the screenshot level and erode the cost savings. The absence of any stated limitations (the classifier found none) leaves this selection bias unacknowledged.

The reader's weakest assumption correctly identifies the sufficiency and representativeness of the accessibility-subtree diff as the load-bearing condition. My stress test agrees: all major quantitative claims (708x bytes, 47.3x cloud calls, one-episode recovery, matching fresh experience) are downstream of the diff repairing most failing steps at low cost. The paper gives no evidence that the diff geometry generalizes beyond the nine curated pairs, and its own data show the diff is not the knee for two of six apps. The concrete test I propose directly measures the repair rate and cost on a held-out set, which would settle whether the 708x factor is an artifact of selection. I do not see an error that would warrant rejection; the absence of a limitations section and lack of released artifacts are addressable conditions, consistent with the reader's CONDITIONAL verdict. Thus I keep the verdict unchanged.

The referee's report

If the reported results are accepted, REBASE is a genuinely useful system contribution. The problem it addresses is concrete and, to the best of the paper's positioning, under-served: experience coherence across app updates is discussed in prior work mainly as a limitation. The paper includes several strengths: a real-device implementation, an automated AndroidWorld evaluator rather than model-judged success, confidence intervals on the headline success gain, an ablation table isolating the update clock and failure clock, blind pairs that test behavior when the cloud cannot obtain the new version, and a trace-driven fleet experiment for the deduplication claim. The main caveat is external validity: the version-pair selection rule and the fixed accessibility-subtree diff geometry mean the headline byte savings are measured on a testbed that is deliberately enriched for the exact repair signal REBASE uses. That issue is central enough to require additional analysis before the broader claims can be accepted, but it does not undermine the soundness of the measurements as reported on the stated testbed.

Pith's reading of the paper (a machine's summary, not the authors')

A phone GUI agent reuses 'experience': action paths recorded on earlier runs, consolidated in the cloud, and cached on the device. When an app updates, part of that experience goes silently wrong — on nine real version pairs, stale experience succeeds 10.1 points less often than no experience at all, and a broken step stays walkable in 62.2% of executions, so the agent does not notice the mismatch until roughly three steps later. The paper's claim is that a two-clock protocol — the cloud replays its copy on the new version while an updated device verifies every recorded step before executing it, and a failing step triggers an escalating ladder of evidence — restores stale experience to the success rate of freshly recorded experience within a single episode. The load-bearing observation behind the cost saving is that an accessibility-subtree diff of only 8.0 KB repairs 79.1% of failing steps, where a 298 KB screenshot adds only about 9.4 points more.

What carries the argument. The central object is the accessibility-subtree diff: a small (median 8.0 KB) record of what changed in the accessibility tree around the expected element, rooted at that element's grandparent and extending three levels with one sibling hop. It is the 'knee' of the evidence ladder — it repairs 79.1% of failing steps, where a fingerprint alone repairs only 27.4% and a 298 KB screenshot adds only about 9.4 points more at a median 36.6x the bytes per step. Around it sits the two-clock protocol: the update clock (cloud replay plus on-device verification $\nu_k$, which checks in two factors that the step's element exists on the new screen and that the resulting screen matches the recorded successor) and the failure clock (staged evidence $\xi_0 < \xi_1 < \xi_2 < \xi_3$ escalating until a verified patch $\pi$ is re-derived and committed under the version key). Verification is deliberately model-free on the device — the fingerprint relation $f \simeq f'$ tests role, hash, normalized label edit distance, and box overlap — so passing steps cost a lookup rather than a small-model call, which accounts for most of the 2.5x device-energy saving.

Core claim. REBASE treats an app update as a coherence event between the device copy and the cloud copy of the agent's experience. On the update clock, the cloud installs the new version on a virtual device, replays its copy until the first failing step, and re-derives an executable patch; on the failure clock, a device that cannot replay verifies each recorded step before executing it and, on mismatch, sends evidence in increasing size — a fingerprint, then an accessibility-subtree diff rooted at the expected element's grandparent and extending three levels with one sibling hop, then a crop, then a screenshot — until the cloud repairs the failed segment and returns a version-keyed patch. Relative to stale experience, REBASE raises Succ@K by 56.1 points (95% CI [51.6, 60.6]) and matches fresh experience recorded on the new version (difference 0.4, CI [−2.1, 3.0]), recovering within the first episode for 91.1% of cloud-visible units, at 708x fewer bytes and 47.3x fewer cloud calls than screenshot-to-cloud takeover, and 2.5x less device energy than running without experience. Because every repair is keyed by $(\alpha, v', f)$ — app, version, fingerprint — one cloud re-derivation serves every device that breaks on the same fingerprint, keeping re-derivations near the count of update-broken steps (about 42) rather than scaling with fleet size (80.0x fewer in a trace-driven 100-device fleet).

The premise it rests on. The load-bearing premise is that app-update breakage is reachable by the accessibility-subtree diff — rooted at the expected element's grandparent, three levels deep with one sibling hop — which repairs 79.1% of failing steps here; if real-world updates break steps this diff geometry cannot see (changes buried deeper than three levels, or screens with no accessibility tree), the ladder degrades toward the screenshot level and the 708x byte reduction with it.

If the paper is right (Pith's inferences, not the authors')

Stale experience is worse than none: an agent carrying experience from the previous version succeeds 10.1 points less often than one carrying none, so a version change should be treated as an invalidation signal rather than a non-event.

A device–cloud protocol can fully recover update-induced performance and energy loss on the evaluated apps, within one episode for most cloud-visible units (91.1%), and within a few episodes even when the cloud cannot obtain the new release (68.9% of blind units).

The evidence ladder lets the link carry a description of what changed rather than raw pixels, cutting uplink from 2912 KB per episode (screenshot takeover) to 2.5 KB, and it stays inside the fingerprint or diff level for 79.4% of ladder events.

Version-keyed patches turn a fleet-wide update into a bounded repair problem: re-derivations stay near the number of update-broken fingerprints instead of growing with the number of devices.

Because verification runs without a model call, protected episodes use 2.5x less device energy than running with no experience, and the median 17 J cost of a coherence event is recovered within 0.02 of an episode.

If the 79.1% diff-repair rate holds beyond the nine selected version pairs, the same staged-evidence ladder could repair other 'stale state' breakdowns — config changes, remote content updates, or A/B experiments that mutate the accessibility tree without a version bump — where the device-side verification cost is already paid.

The crop fallback for tree-less screens (canvases, opaque web views) means the byte savings are conditional on an accessibility tree being present; a deployment on tree-poor apps would degrade toward the screenshot level, so the 708x figure is best read as an upper bound for tree-rich apps.

Version-keyed dedup points to a fleet-wide economy the paper leaves implicit: if the broken-fingerprint set is small and slowly changing, a cloud service could precompute and push the patch set at release time (the 'push set' $U$), eliminating most device-originated events entirely.

The measured harm of stale experience suggests a cheap, testable floor for practice: any agent that cannot verify a recorded step before executing it should probably run model-only rather than reuse a stale path, a baseline worth adding alongside re-explore.

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

Every value above, in the review's public record (JSON)

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.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

A graphical user interface (GUI) agent that ships on a phone runs a small model and reuses experience: action paths recorded on earlier runs and cached from the cloud. When an app updates, part of this experience becomes silently wrong. On nine real app version pairs, an agent carrying experience recorded on the old version succeeds 10.1 points less often than one carrying none, and it does not notice the mismatch until 3 steps after acting on it. We propose REBASE, a protocol that keeps the device copy and the cloud copy of the experience coherent across updates. When a version changes, the cloud replays its copy on the new version and the device verifies each recorded step before executing it; when a step still fails, the device sends the cloud evidence in increasing size, starting from an accessibility-subtree diff, until the cloud re-derives an executable patch, keyed by version so that one repair serves the whole fleet. On two Jetson devices, REBASE restores the success rate of stale experience to that of fresh experience recorded on the new version within one episode, at 708x fewer bytes and 47.3x fewer cloud calls than uploading screenshots to the cloud, and 2.5x less device energy than running without experience.

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 0.96

“On two Jetson devices, REBASE restores the success rate of stale experience to that of fresh experience recorded on the new version within one episode, at 708x fewer bytes and 47.3x fewer cloud calls than uploading screenshots to the cloud, and 2.5x less device energy than running without experience.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.32 · human participants 0.05 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.94 · mentions code or data availability 0.05

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.

SentenceRoleThe authors' own workStates the main conclusionStrength
A graphical user interface (GUI) agent that ships on a phone runs a small model and reuses experience: action paths recorded on earlier runs and cached from the cloud. background 0.99 0.22 0.03 asserted 0.99
When an app updates, part of this experience becomes silently wrong. background 0.84 0.28 0.04 asserted 0.99
On nine real app version pairs, an agent carrying experience recorded on the old version succeeds 10.1 points less often than one carrying none, and it does not notice the mismatch until 3 steps after acting on it. result 0.92 0.91 0.10 asserted 1.00
We propose REBASE, a protocol that keeps the device copy and the cloud copy of the experience coherent across updates. contribution 1.00 0.60 0.16 asserted 0.92
When a version changes, the cloud replays its copy on the new version and the device verifies each recorded step before executing it; when a step still fails, the device sends the cloud evidence in increasing size, starting from an accessibility-subtree diff, until the cloud re-derives an executable patch, keyed by version so that one repair serves the whole fleet. method 0.97 0.79 0.14 asserted 0.98
On two Jetson devices, REBASE restores the success rate of stale experience to that of fresh experience recorded on the new version within one episode, at 708x fewer bytes and 47.3x fewer cloud calls than uploading screenshots to the cloud, and 2.5x less device energy than running without experience. result 1.00 0.95 0.70 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00

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, 5,119 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 (5)
Figure 1 of the paper
Figure 1. Stale experience fails silently, while accessibility-subtree diffs enable most repairs. Left: Success on new versions of nine apps with no experience or old-version experience; badges report mean loss by update type. Right: Share of failing steps repaired from five evidence levels across six cloud-visible apps. The diff marks the knee in four of six apps. Three seeds. ence comprises executable paths bound to accessibility-tree elements [21]. Each step uses a fingerprint of the element’s role, label, and position. The agent walks a path by executing its steps in order. The cloud…
Figure 2 of the paper
Figure 2. REBASE overview. The update clock drives verification on the device, which checks each recorded step before and after acting, and release-time replay on the cloud (A), which precomputes patches. The failure clock drives evidence-driven repair (B): the lightest sufficient evidence re-derives the failed segment, which is tested and returned under its version key for reuse. 0.1. Let S denote a screen’s fingerprint set. We write f ≃ S when f matches an element of S and use the same relation for screens whose sets overlap by at least θs = 0.5. Verification νk is one when the step-k…
Figure 3 of the paper
Figure 3. The REBASE testbed: (A) two screens of a walked path, (B) the devices, and (C) the topology. IV. IMPLEMENTATION AND TESTBED Two NVIDIA Jetson AGX Orin 64 GB boards are the de￾vices (
Figure 4 of the paper
Figure 4. Link tiers: three pairs, five tasks, K = 5, and three seeds (bars: mean ± standard deviation). Succ@K against (a) uplink bandwidth at 80 ms and (b) round-trip time at 10 Mbps; (c) episode latency against bandwidth; (d) episode latency against transferred kilobytes, with one point per episode and a 0.5 KB floor for zero-byte episodes. 3 10 30 100 300 Fleet size N 10 102 103 104 105 Re-derivations Version-key dedup No dedup Real devices (N = 3) Model: |D|, N |D| (a) Re-derivations vs. fleet size No exp.Stale Re-explore Shot to cloud Trajectory Cloud-only REBASE 0 2 4 Energy (kJ) (b)…
Figure 5 of the paper
Figure 5. Fleets and energy. (a) Cloud re-derivations vs. fleet size

Figures are the authors', reproduced from the paper.

References (48)
  1. [1] B. Wu, Z. Ding, J. Huang, and Y . Zhao, “Forget to Improve: On-Device LLM-Agent Continual Learning via Budget-Curated Memory,” arXiv preprint arXiv:2606.25115, 2026 on PithPith's reviewarXiv
  2. [2] B. Wu, Z. Ding, and J. Huang, “ERRAND: Budgeted Maintenance of Agent Memory,” arXiv preprint arXiv:2609.29545, 2026 on PithPith's reviewarXiv
  3. [3] H. Zhouet al., “MAI-UI Technical Report: Real-World Centric Foun- dation GUI Agents,” arXiv preprint arXiv:2512.22047, 2025 on PitharXiv
  4. [4] Y . Jiang and C. Huang, “OpenPhone: Mobile Agentic Foundation Models,” inFindings of the Association for Computational Linguistics: ACL 2026, 2026, pp. 30 362–30 380 on Pith
  5. [5] S. Li, Z. Zuo, H. Wang, J. Chen, Z. Jin, and R. LI, “Administrative Decentralization in Edge-Cloud Multi-Agent for Mobile Automation,” arXiv preprint arXiv:2604.07767, 2026 on PithPith's reviewarXiv
  6. [6] G. Fan, C. Niu, C. Lyu, F. Wu, and G. Chen, “CORE: Reducing UI Exposure in Mobile Agents via Collaboration Between Cloud and Local LLMs,” inAdvances in Neural Information Processing Systems, 2025 on Pith
  7. [7] Z. Xie, Y . Xu, H. Xu, Y . Liao, and Z. Yao, “A Novel Hat-Shaped Device- Cloud Collaborative Inference Framework for Large Language Models,” inProceedings of the IEEE International Conference on Computer Communications, 2026 on Pith
  8. [8] Z. Wanget al., “PPAI: Enabling Personalized LLM Agent Interoper- ability for Collaborative Edge Intelligence,” inProceedings of the IEEE International Conference on Computer Communications, 2026 on Pith
  9. [9] Y . Zhanget al., “Hera: Learning Long-Horizon Coordination for Device- Cloud Collaborative LLM Agents,” arXiv preprint arXiv:2605.24598, 2026 on PithPith's reviewarXiv
  10. [10] Z. Fanget al., “R-ACP: Real-Time Adaptive Collaborative Perception Leveraging Robust Task-Oriented Communications,”IEEE Journal on Selected Areas in Communications, vol. 43, no. 12, pp. 4215–4230, 2025 on Pith
  11. [11] Z. Qin, S. Yue, X. Hua, Y . Fu, and J. Ren, “Executable Agentic Memory for GUI Agent,” inProceedings of the International Conference on Machine Learning, 2026 on Pith
  12. [12] H. Miet al., “Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution,” inProceedings of the International Conference on Machine Learning, 2026 on Pith
  13. [13] T. Men, Z. Jin, P. Cao, Y . Chen, K. Liu, and J. Zhao, “Empowering GUI Agents via Autonomous Experience Exploration and Hindsight Experi- ence Utilization for Task Planning,” inProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: L... on Pith
  14. [14] Y . Wang, L. Yang, Z. Liu, A. Kumar, and C. Maple, “CoMIC: Collabo- rative Memory and Insights Circulation for Long-Horizon LLM Agents in Cloud-Edge Systems,” arXiv preprint arXiv:2606.00756, 2026 on PithPith's reviewarXiv
  15. [15] J. Saad-Falconet al., “OpenJarvis: Personal AI, On Personal Devices,” arXiv preprint arXiv:2605.17172, 2026 on PithPith's reviewarXiv
  16. [16] Healenium, mabl, and Testim, “Self-healing locators in test automation,” https://healenium.io/docs/how healenium works; https://www.mabl. com/auto-healing-tests; https://www.testim.io/blog/announcing-auto- improving-smart-locators-dare-we-say-genius-locators/, accessed: 2026-09-09 on Pith
  17. [17] S. Caoet al., “Comprehensive Semantic Repair of Obsolete GUI Test Scripts for Mobile Applications,” inProceedings of the IEEE/ACM 46th International Conference on Software Engineering, 2024 on Pith
  18. [18] S. Cao, M. Pan, and X. Li, “TUSR: A Test Unit-Based Framework for Repairing Obsolete GUI Test Scripts,”Proceedings of the ACM on Software Engineering, vol. 3, no. FSE, pp. 4714–4736, 2026 on Pith
  19. [19] B. Wu, Z. Ding, and J. Huang, “A Review of Continual Learning in Edge AI,”IEEE Transactions on Network Science and Engineering, vol. 13, pp. 6571–6588, 2026 on Pith
  20. [20] B. Wu and J. Huang, “Lifecycle-Aware Federated Continual Learning in Mobile Autonomous Systems,” arXiv preprint arXiv:2604.20745, 2026 on PithPith's reviewarXiv
show 28 more
  1. [21] Android Developers, “Create an accessibility service,” https://developer. android.com/guide/topics/ui/accessibility/service, 2026, accessed: 2026- 09-09 on Pith
  2. [22] B. Wu and J. Huang, “CrystalMem: Elastic Memory for Self- Evolving LLM Agents via Knowledge Crystallization,” arXiv preprint arXiv:2608.00303, 2026 on PithPith's reviewarXiv
  3. [23] Z. Fang, Y . Guo, Y . Zhang, H. An, W. Ding, and Y . Fang, “Shared Spatial Memory Through Predictive Coding,” arXiv preprint arXiv:2511.04235, 2025 on PitharXiv
  4. [24] B. Wu, J. Huang, Q. Duan, L. Dong, and Z. Cai, “Enhancing Vehicular Platooning With Wireless Federated Learning: A Resource-Aware Con- trol Framework,”IEEE/ACM Transactions on Networking, vol. 34, pp. 1479–1494, 2026 on Pith
  5. [25] L. Dong, J. Huang, and R. W. Heath, “Transformer-Based Dynamic Resource Allocation for Multi-Carrier NOMA Systems,”IEEE Transac- tions on Cognitive Communications and Networking, vol. 12, pp. 4926– 4941, 2026 on Pith
  6. [26] C.-C. Xing, Z. Ding, and J. Huang, “A Stochastic Geometry-Based Analysis of SWIPT-Assisted Underlaid Device-to-Device Energy Har- vesting,”ACM SIGAPP Applied Computing Review, vol. 25, no. 4, pp. 18–34, 2025 on Pith
  7. [27] B. Wu, Z. Cai, W. Wu, and X. Yin, “AoI-Aware Resource Management for Smart Health via Deep Reinforcement Learning,”IEEE Access, vol. 11, pp. 81 180–81 195, 2023 on Pith
  8. [28] Z. Ding, B. Wu, and J. Huang, “SCALE: Sensitivity-Aware Federated Unlearning with Information Freshness Optimization for Mobile Edge Computing,” inProceedings of the IEEE International Conference on Distributed Computing Systems, 2026, pp. 1476–1486 on Pith
  9. [29] B. Wu, J. Huang, and Y . Zhao, “From Alpha to Omega: Lifecycle- Aware Forgetting Defense in Federated Continual Learning for Planetary Exploration,” inProceedings of the IEEE International Conference on Distributed Computing Systems, 2026, pp. 1–11 on Pith
  10. [30] B. Wu, Z. Ding, and J. Huang, “Exposing and Resolving Spurious Iso- lation in Federated Multimodal Continual Learning,”IEEE Transactions on Multimedia, 2026 on Pith
  11. [31] B. Wu and W. Wu, “Model-Free Cooperative Optimal Output Regulation for Linear Discrete-Time Multi-Agent Systems Using Reinforcement Learning,”Mathematical Problems in Engineering, vol. 2023, no. 1, p. 6350647, 2023 on Pith
  12. [32] Z. Ding, J. Huang, and J. Qi, “Learning to Defend: A Multi-Agent Reinforcement Learning Framework for Stackelberg Security Game in Mobile Edge Computing,” inProceedings of the International Confer- ence on Computing, Networking and Communications, 2026, pp. 769– 774 on Pith
  13. [33] J. Huang, B. Wu, Z. Ding, and L. Ostigaard, “Reinforcement Learning- Based Energy-Aware Coverage Path Planning for Precision Agriculture,” inProceedings of the International Conference on Research in Adaptive and Convergent Systems, 2025, pp. 1–8 on Pith
  14. [34] J. Huang, B. Wu, Q. Duan, L. Dong, and S. Yu, “A Fast UA V Tra- jectory Planning Framework in RIS-Assisted Communication Systems With Accelerated Learning via Multithreading and Federating,”IEEE Transactions on Mobile Computing, vol. 24, no. 8, pp. 6870–6885, 2025 on Pith
  15. [35] B. Wu, J. Huang, and Q. Duan, “FedTD3: An Accelerated Learning Approach for UA V Trajectory Planning,” inProceedings of the Interna- tional Conference on Wireless Artificial Intelligent Computing Systems and Applications, 2025, pp. 13–24 on Pith
  16. [36] B. Wu, J. Huang, and Q. Duan, “Real-Time Intelligent Healthcare Enabled by Federated Digital Twins With AoI Optimization,”IEEE Network, vol. 40, no. 2, pp. 184–191, 2026 on Pith
  17. [37] Z. Ding, B. Wu, J. Huang, and S. Mao, “Application-Aware Twin-in- the-Loop Planning for Federated Split Learning over Wireless Edge Networks,” arXiv preprint arXiv:2604.26105, 2026 on PithPith's reviewarXiv
  18. [38] Z. Fang, S. Hu, J. Wang, Y . Deng, X. Chen, and Y . Fang, “Prioritized Information Bottleneck Theoretic Framework With Distributed Online Learning for Edge Video Analytics,”IEEE/ACM Transactions on Net- working, vol. 33, no. 3, pp. 1203–1219, 2025 on Pith
  19. [39] Z. Ding, B. Wu, and J. Huang, “EASE: Federated Multimodal Un- learning via Entanglement-Aware Anchor Closure,” arXiv preprint arXiv:2605.00733, 2026 on PithPith's reviewarXiv
  20. [40] Z. Ding and J. Huang, “Toward Trustworthy Federated Unlearning for Mobile Autonomous Systems,”IEEE Network, 2026 on Pith
  21. [41] D. Y . Tchalla, B. Wu, J. Huang, S. Gu, and Q. Duan, “FedSceneX: Time- to-Target Orchestration for Same-Scene Multimodal Federated Edge Learning,” arXiv preprint arXiv:2608.07730, 2026 on PithPith's reviewarXiv
  22. [42] Z. Ding, J. Huang, Q. Duan, C. Zhang, Y . Zhao, and S. Gu, “A Dual-Level Game-Theoretic Approach for Collaborative Learning in UA V-Assisted Heterogeneous Vehicle Networks,” inProceedings of the IEEE International Performance, Computing, and Communications Conference, 2025, pp. 1–8 on Pith
  23. [43] B. Wu, Z. Ding, and J. Huang, “RELIEF: Turning Missing Modalities into Training Acceleration for Federated Learning on Heterogeneous IoT Edge,”IEEE Internet of Things Journal, 2026 on Pith
  24. [44] D. Pan, B.-N. Wu, Y .-L. Sun, and Y .-P. Xu, “A Fault-Tolerant and Energy-Efficient Design of a Network Switch Based on a Quantum- Based Nano-Communication Technique,”Sustainable Computing: In- formatics and Systems, vol. 37, p. 100827, 2023 on Pith
  25. [45] S. Baiet al., “Qwen3-VL Technical Report,” arXiv preprint arXiv:2511.21631, 2025 on PithPith's reviewarXiv
  26. [46] C. Rawleset al., “AndroidWorld: A Dynamic Benchmarking Environ- ment for Autonomous Agents,” inInternational Conference on Learning Representations, 2025 on Pith
  27. [47] B. Wu, J. Huang, and S. Yu, ““X of Information” Continuum: A Survey on AI-Driven Multi-Dimensional Metrics for Next-Generation Net- worked Systems,”IEEE Communications Surveys & Tutorials, vol. 28, pp. 5307–5344, 2026 on Pith
  28. [48] Q. Duan, J. Huang, S. Hu, R. Deng, Z. Lu, and S. Yu, “Combining Fed- erated Learning and Edge Computing Toward Ubiquitous Intelligence in 6G Network: Challenges, Recent Advances, and Future Directions,” IEEE Communications Surveys & Tutorials, vol. 25, no. 4, pp. 2892– 2950, 2023 on Pith
Cited by 3 Pith papers

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.

No comments yet.

Sign in to comment