简体中文 · English
Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents¶
L1 core mechanism diagnostic. Uses only public observations to demonstrate bounded skill revision. No MobileWorld, AndroidWorld or OSWorld run.
Paper information¶
| Field | Details |
|---|---|
| Paper | Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents |
| Institution | Zhejiang University (first author affiliation) |
| First public release | 2026-09-15 (arXiv v1) |
| Original source code | Yes: ZJU-REAL/EvoSkill-GUI |
| Adapter/method | evoskill-gui |
| Local reproduction code | src/auto_research/agent_research/latest_20260919.py |
Original paper summary¶
Background and main changes¶
Split GUI skills into editable components. After execution fails, a critic with isolated information reflects on the failure and revises only the responsible component. Verified skills can be reused on other tasks.
flowchart LR
I[Public input and current state] --> M[Bounded skill revision]
M --> A[Auditable intermediates]
A --> O[Output or state update]Key figure from the paper¶
Figure 2: Core architecture, modules and connections. From the paper; copyright belongs to the authors. Click for the source.
Core mechanism¶
The reference kernel retains defining gates, weights, state transitions or scheduling rules and records intermediates in metrics. See implementation functions and test invariants for variable and formula mappings.
Offline and online results¶
The paper's online, benchmark, training-efficiency and inference-speed results are author results only. The local three-seed artifact checks the core mechanism, shapes and state invariants, not paper-level scores.
Scope¶
Local evidence comes only from public observations and demonstrates bounded skill revision. MobileWorld, AndroidWorld and OSWorld were not run.