- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper presents EASE, a mobile GUI agent that uses a lightweight executor for interaction and a stronger planner for long tasks. The executor keeps a short history summary as it works. The authors introduce TAPO to train action and summary generation together, and build Skill Memory from execution traces to help the planner choose suitable subgoals. Results on GUI-Odyssey and MemGUI-Bench show improvements in performance and efficiency.
Critical strengths
- Having the executor maintain its own summary is a practical way to keep the context manageable without a separate summarization call.
- TAPO gives actions and summaries separate training signals. This is a sensible way to account for their different roles.
- The experiments cover both performance and cost, with useful ablations of the main components.
Critical weaknesses
-
The planner–executor setup and history summarization are established ideas. The paper could explain more clearly what TAPO and the skill-construction method add beyond prior work.
-
The full system is evaluated online, but the direct comparison between TAPO, GRPO, and GDPO is limited to offline action prediction. It is unclear how much TAPO improves end-to-end task success.
-
Repeated summarization may lose important information. The paper provides REMEMBER and task-level storage, but does not closely examine when these mechanisms fail.
Questions for authors
-
Does TAPO outperform GRPO and GDPO in online evaluation when the planner, skill memory, and interaction budget are kept the same?
-
How often do missing information or summary errors cause task failures, even with REMEMBER and task-level storage?
-
How much of the cost reduction comes from fewer planner calls versus lower token usage per call? A token breakdown would be helpful.
5 comments
This is Part 2/2 of our response, addressing Weakness 2 / Question 1 and Weakness 3 / Question 2.
Reviewer (Weakness 2 & Question 1): W2: The full system is evaluated online, but the direct comparison between TAPO, GRPO, and GDPO is limited to offline action prediction. It is unclear how much TAPO improves end-to-end task success.
Q1: Does TAPO outperform GRPO and GDPO in online evaluation when the planner, skill memory, and interaction budget are kept the same?
Thank you for highlighting this limitation. We conducted additional online evaluations on MemGUI-Bench, keeping the planner, Skill Memory, interaction budget, and all other settings identical. Only the executor differs, using models trained with GRPO, GDPO, or TAPO.
| Executor training | pass@1 (%) | pass@3 (%) |
|---|---|---|
| GRPO | 11.7 | 25.8 |
| GDPO | 13.3 | 24.2 |
| TAPO (EASE-Agent) | 19.5 | 31.3 |
TAPO outperforms GRPO and GDPO by 7.8 and 6.2 percentage points in pass@1, and 5.5 and 7.1 points in pass@3, respectively. These results demonstrate that TAPO's improvements extend beyond offline action prediction to end-to-end task completion.
Reviewer (Weakness 3 & Question 2): W3: Repeated summarization may lose important information. The paper provides REMEMBER and task-level storage, but does not closely examine when these mechanisms fail.
Q2: How often do missing information or summary errors cause task failures, even with REMEMBER and task-level storage?
We agree that repeated summarization may introduce errors despite the use of REMEMBER and task-level storage. To better understand these limitations, we analyzed 286 failed attempts from EASE-Agent's online evaluation (19.5% pass@1 and 31.3% pass@3, Table 2).
Each attempt was assigned to one primary failure category based on the earliest identifiable failure-related error that was not effectively corrected. We further examined whether errors in the history summary or REMEMBER mechanism contributed to each failure.
| Outcome | Primary failure category | Failures (%) | Incorrect summary | Progress lost | Not remembered | Incorrectly remembered |
|---|---|---|---|---|---|---|
| Reported completion | Incorrect information used | 5 (1.7%) | 0 | 0 | 2 | 3 |
| Reported completion | Required operations omitted | 29 (10.1%) | 26 | 1 | 0 | 0 |
| Did not complete | Stalled during information retrieval | 103 (36.0%) | 10 | 0 | 2 | 13 |
| Did not complete | Stalled during GUI operations | 126 (44.1%) | 4 | 0 | 0 | 0 |
| Did not complete | Unreported app/login blockage | 23 (8.0%) | 0 | 0 | 0 | 0 |
| Total | 286 (100%) | 40 | 1 | 4 | 16 |
Primary failure categories are mutually exclusive, while memory-error annotations may overlap. We count a memory error only when the execution trace indicates that it contributed to failure. All percentages are calculated over the 286 failed attempts.
Overall, incorrect summaries contributed to 40 failures (14.0%), primarily involving falsely reporting incomplete operations as completed, misrepresenting execution progress, or incorrectly recording the outcomes of GUI interactions. In contrast, loss of progress was identified in only one case (0.3%). Incorrectly remembered information contributed to 16 failures (5.6%), leading to errors in subsequent task execution.
The two mechanisms serve complementary roles: history summaries provide a compact overview of execution progress, rather than preserving every interaction detail, while REMEMBER explicitly retains task-critical information that might otherwise be lost through repeated summarization. Nevertheless, our analysis shows that inaccurate progress tracking and missing or incorrect retained information can still contribute to task failures. We will incorporate these findings and discuss these limitations in the revised paper.
We sincerely appreciate the reviewer's constructive suggestions. We hope these additional experiments and analyses help clarify our methodological contributions and the limitations of our approach.
Thank you for your response. I have another question about Table 2: why does the model perform better on tasks involving 4 or 3 apps than on those involving 2 apps? Intuitively, tasks involving more apps would seem more difficult and might be expected to yield lower performance.
I have a similar question about task difficulty: why is performance better on hard tasks than on medium tasks? Could you explain these results?
We sincerely thank the reviewer for the thoughtful and constructive feedback, as well as the positive assessment of our work. Due to the comment length limit, we divide our response into two parts. This first part addresses Weakness 1 and Question 3.
We agree that hierarchical planning and history summarization are established ideas. Our contributions focus on addressing two limitations in their integration.
Regarding TAPO, jointly training action generation and history summarization introduces cross-objective interference. GRPO assigns a shared advantage to both outputs, while GDPO separately normalizes rewards but still combines their advantages. TAPO instead assigns objective-specific advantages to the corresponding action and summary tokens, providing more targeted training signals. As shown in Table 5, TAPO achieves 64.9% SR, outperforming both GRPO and GDPO by 1.9 percentage points, without requiring additional rollouts or training stages.
Regarding Skill Memory, our contribution lies primarily in its role in executor-calibrated planning. While prior work primarily provides reusable procedural and operational guidance for GUI tasks, our Skill Memory guides task decomposition according to the executor's observed capabilities, avoiding overly coarse subgoals and unnecessary planner interventions. We construct Skill Memory by comparing execution outcomes across different planning granularities, rather than relying on naturally collected trajectories. As shown in Table 6, this multi-granularity construction improves pass@3 from 28.1% to 31.3%.
Thank you for this suggestion. We provide a token-level breakdown to clarify the sources of cost reduction. M3A relies entirely on Gemini-2.5-Pro, while EASE-Agent-Pro/Max use the same model as the planner and delegate frequent GUI interactions to lower-cost executors. The following results are measured under pass@3.
For M3A, planner calls refer to all Gemini-2.5-Pro invocations.
Compared with M3A, EASE-Pro and EASE-Max reduce expensive Gemini-2.5-Pro calls by 93.6% and 92.8%, respectively, while reducing input tokens per strong-model call by 27.7% and 39.8%. Although output tokens per call increase, these results indicate that reducing expensive model invocations, rather than merely reducing tokens per call, is the primary source of efficiency gains.
This improvement comes from allocating tasks according to model capabilities: unlike M3A, which repeatedly invokes Gemini-2.5-Pro during GUI interaction, EASE delegates most GUI operations to lower-cost executors and invokes the strong planner only for task decomposition, subgoal verification, and replanning. Self-Maintained State enables the executor to maintain execution progress independently, while executor-calibrated Skill Memory further reduces unnecessary planner interventions (Figure 5). As a result, EASE-Pro and EASE-Max reduce API cost per attempt by 73.7% and 69.1%, respectively, while achieving higher pass@3 (50.0% and 75.0% versus M3A's 47.7%).
Our responses to Weakness 2 / Question 1 and Weakness 3 / Question 2 are provided in a separate comment (Part 2/2).