- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
CoAdapt-GUI adapts both a transferable workflow context and a GUI policy to a new application. It evolves context variants with rollout feedback while updating LoRA policy parameters using task-context-matched groups. The paper evaluates unseen-app adaptation in Android-Generalization and AndroidWorld Plus, with the latter separating adaptation and evaluation task templates. Joint adaptation improves reported success rates over static transfer and either adaptation component alone.
Critical strengths
-
The method addresses a practical mismatch between reusable workflow knowledge and the behavior of a policy executing that knowledge. The immutable source/mutable target distinction and task-context-matched update groups are sensible design choices, and the implementation is described in sufficient detail to understand their intended role.
-
The experiments include relevant component baselines and a stronger template-disjoint evaluation in AndroidWorld Plus. Joint adaptation reaches 52.57% success versus 47.24% for context-only and 40.38% for policy-only adaptation; the shared/novel-category breakdown makes the remaining transfer difficulty visible. Freezing adaptation before evaluation is an important strength.
Critical weaknesses
-
The evidence establishes a useful joint system more clearly than it establishes the proposed mechanisms. Removing all context or policy adaptation does not isolate the transfer constraints, the context-matched credit assignment, or the population/rating mechanism. Appendix E primarily compares the main configurations rather than ablating those choices. Since joint prompt/weight adaptation is already discussed in the related work, identifying which GUI-specific design contributes beyond simply running both components is central to the methodological contribution.
-
The adaptation and selection comparisons need clearer accounting. Appendix B.3 states that nominal update counts do not constitute matched compute, while joint adaptation includes context generation/reflection and candidate evaluation. The reported variation is also insufficiently explicit about independently repeated adaptation runs versus repeated evaluation of adapted checkpoints. Finally, Appendix C.3 retains historical context ratings across policy changes without final-policy reranking; this leaves the selected context's compatibility with the final policy empirically unresolved. These limitations make the size and reproducibility of the claimed co-adaptation advantage difficult to assess.
Questions for authors
-
Under the same environment-interaction budget, what happens when task-context grouping is replaced by task-only grouping, and when the workflow transfer constraints are removed? A focused ablation of the claimed mechanisms would materially affect my assessment.
-
Please report realized environment calls, optimizer updates, reflection/generation calls, and adaptation seeds for each main baseline. How much of the joint gain remains across independent adaptation runs?
-
Does reevaluating/reranking candidate contexts under the final policy materially change the chosen context or test performance?
Official Comment by Reviewer WNov
The rationale for using UI-TARS-7B on AndroidWorld-Generalization but Qwen3-VL-8B-Instruct on AndroidWorld Plus is unclear. Please justify this choice and evaluate both backbones on the same benchmark to assess whether the improvements are consistent across models.
Why not just use "Base Policy" as the model name?
Unlike Table 2, some models in Table 1 span two rows. This is the reason for my question; if the authors do not consider this an issue—or if it was intentional—then please feel free to disregard this point.