Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents
Bofan Chen, Boxuan Zhang, Fei Tang, Zhengxi Lu, Yong Du, Tongbo Chen, Weiming Lu, Jun Xiao, et al.
cs.LG · cs.AI
Quick verdict
Training-free GUI skill self-evolution with consistent gains and strong ablations, but lacks significance tests and head-to-head rival baselines.
Summary
The paper proposes EvoSkill-GUI, a training-free framework that treats GUI-agent skills as living, editable artifacts rather than static prompts. Each skill is a multi-file package (retrieval metadata, executable plan, backup localization, recovery rules, accessibility utilities, failure examples) manipulated through a restricted file-tool interface. A reflect-revise-reuse loop lets the executor apply in-rollout edits, then—after failed rollouts—serve as its own information-isolated critic before revising specific skill files. On MobileWorld, AndroidWorld, and OSWorld the framework improves several backbones (e.g., Qwen3.6-Plus 53.3%→69.5% on MobileWorld; Qwen3-VL-8B 23.8%→34.3% on OSWorld), with ablations attributing gains to information isolation (−8.57 when removed) and instant revision (−6.66). Analyses cover held-out transfer (73.3% vs 40.0% pass@3), critic-audit accuracy (80.4%), repairability by failure type, and token-budget fairness versus repeated sampling. Main limitations: no significance testing or multi-seed variance behind headline gains, no head-to-head comparison with competing self-evolving/memory baselines, retrieval hyperparameters tuned on evaluated tasks, reuse statistics measured on parameterized variants of the same task families, and no verifier against harmful skill edits—most of which the authors themselves acknowledge.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- GUI agents need skills that are revised from execution feedback at deployment time rather than static pre-built skills, and this can be done without any training.
- Structured multi-file skill packages enable targeted revision and outperform monolithic single-file skills (+2.85 points on MobileWorld).
- A single backbone acting as both executor and information-isolated critic yields better diagnoses than an unisolated setup (+8.57 points vs w/o isolation).
- Evolved skill libraries transfer to related and held-out tasks (73.3% vs 40.0% for pass@3 on a held-out MobileWorld split) and beat repeated sampling at comparable token cost (69.5% vs 62.8%).
- Across MobileWorld, AndroidWorld, and OSWorld the framework consistently improves multiple base models, with maximum gains of +16.2, +6.0, and +10.5 points respectively.
Strengths
Core mechanisms are cleanly isolated by ablations
Table 4 (MobileWorld, Qwen3.6-Plus) shows removing information isolation drops SR 69.52%→60.95% (−8.57) and removing instant revision drops it to 62.86% (−6.66); Figure 5a ablates package components (w/o planning docs −3.81, w/o failure examples −2.85), directly supporting the two central design claims.
Broad evaluation across platforms and model classes
Sec. 4 evaluates three benchmarks (mobile MobileWorld/AndroidWorld, desktop OSWorld) over five backbones spanning closed-weight (Claude-Sonnet-4.6, Qwen3.6-Plus), open general (Qwen3.6-35B-A3B), and GUI-specialized (GUI-Owl-1.5-8B, MAI-UI-8B) models, showing the gains are not backbone-specific.
Unusually honest self-auditing
Budget-fair and transfer analyses address the obvious confound
Failure-type repairability analysis connects design to evidence
Weaknesses
No statistical rigor behind point deltas
No head-to-head comparison with rival self-evolving/memory systems
Headline gains are per-benchmark maxima
Retrieval hyperparameters tuned on the evaluation distribution
No safeguard against harmful skill edits and oracle-dependent loop control
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EvoSkill-GUI in simple terms?
It is a training-free framework that lets GUI agents (phone/computer automation agents) store each skill as a structured folder of files—plan, backup locators, recovery rules, failure examples—and automatically edit those files after failed executions, so the skill gets better every time it is reused, with no model fine-tuning.
How much does EvoSkill-GUI improve GUI agent performance?
In the paper's experiments it improves every tested backbone across three benchmarks: up to +16.2 points on MobileWorld (Qwen3.6-Plus, 53.3%→69.5%), up to +6.0 on AndroidWorld (seed 42), and up to +10.5 on OSWorld (Qwen3-VL-8B, 23.8%→34.3%). These are maximum per-benchmark gains from single runs without statistical significance testing.
Does EvoSkill-GUI require training or an extra reflection model?
No. The same backbone model plays both executor and critic in separate sessions: the critic is given only the instruction and the observed trajectory (strictly no access to the skill content, the executor's reasoning, or ground-truth answers), then the executor rewrites specific skill files through a restricted tool interface.
Is the code and data for EvoSkill-GUI available?
The paper states code is released at github.com/ZJU-REAL/EvoSkill-GUI and provides full prompts (executor, skill generation, critic, revision) and hyperparameters in Appendix A, though the review could not verify the repository contents.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The metadata abstract is truncated mid-sentence; the review relies on the full-text abstract embedded in the paper body.
- Table 1 extraction is partially garbled (column alignment shifted; the MMSkills baseline row lacks an Overall value), so individual OSWorld domain numbers should be verified against the PDF.
- Figures (1, 2, 3-8) are available only as extracted text and captions; visual content could not be independently verified.
- References dated 2025-2026 (concurrent work) could not be checked against actual sources since no network access was used.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.