- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
TRACK is a framework for acquiring and using application knowledge in mobile GUI agents. It explores applications offline using functional anchors, depth-first search, and transition verification, then produces task paths, element knowledge, and reusable skills.
During execution, this method localizes progress within retrieved paths and selects knowledge for the current decision.
The paper evaluates TRACK on 129 selected tasks from AndroidWorld and DroidTask with two decision models, and reports comparisons, ablations, and integration into other agents.
Critical strengths
-
The paper addresses a practical limitation of retrieval for GUI agents: relevant knowledge can still suggest an inappropriate action, when prerequisites have not been met or earlier steps have already been completed. Linking knowledge use to execution progress is a reasonable way to address this problem.
-
The results provide useful evidence that selective knowledge use can improve execution. On AndroidWorld, this method increases the GLM success rate from 65.12% without the knowledge base to 83.72%. It also outperforms simple similarity-based injection with the same frozen knowledge base on both models. These comparisons support the value of selecting knowledge rather than injecting retrieved records indiscriminately.
Critical weaknesses
- The contribution beyond existing methods is not sufficiently established.
In previous works, GUI-explorer[a] already combines functional anchors, DFS exploration, state restoration, and transition knowledge. Mobile-Agent-RAG[b] retrieves planning and operational knowledge at different levels, with operational retrieval conditioned on the current subtask. TRACK's main distinction therefore appears to be explicit path progress localization and conditional use of three knowledge types. However, Equation (4) and Appendix A.3 do not specify how the compatibility scores, weights, and localization confidence are computed. The supplementary reference code implements ranking and fixed truncation, but does not show the proposed joint localization and conditional routing. Table 3 also removes path knowledge and localization together. The current evidence does not separate the proposed progress mechanism from the benefits of filtering or reducing retrieved context. [a] Xie B, Shao R, Chen G, et al. Gui-explorer: Autonomous exploration and mining of transition-aware knowledge for gui agent[C]//Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025: 5650-5667. [b] Zhou Y, Li J, Zhang Y, et al. Mobile-agent-rag: driving smart multi-agent coordination with contextual knowledge empowerment for long-horizon mobile automation[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2026, 40(35): 29939-29947.
- The evaluation does not yet provide sufficiently reliable evidence for the broader claims.
Section 4.2 states that DroidTask completion depends on the agent's self-judgment, while the experimental settings state that the original evaluators are retained. The original AutoDroid evaluation measures completion against annotated action sequences. It is unclear whether self-judgment only determines termination or also determines the reported success rate. The latter would make actual completion difficult to assess independently. AndroidWorld provides clearer evidence, but includes only 43 tasks. TRACK exceeds GUI-explorer by one successful task with Qwen and three with GLM, without reported variability across runs. These results suggest potential benefits, but do not establish a stable advantage that resolves the uncertainty about the main contribution.
Questions for authors
- Please clarify how the compatibility terms and weights in Equation (4), the localization confidence threshold, and the semantic step completion checks are implemented. Currently it seems not clear how does the supplied reference code relate to the implementation used for the experiments?
- Does the agent's self-judgment in DroidTask determine only when execution stops, or also whether a task is counted as successful? How is actual completion verified independently, and how does this procedure relate to the original evaluator?
- To validate the main distinction, can you compare progress localization with retrieval based on the current UI or subtask while retaining all three knowledge types, the same planner, and comparable context budgets? Limited repeated runs or paired task results for this comparison would help establish whether the benefit is stable. This request concerns validation of the present claims rather than an extension to new benchmarks.