Literature
Display
← All papers

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 · Yueting Zhuang · Yongliang Shen

· arXiv (preprint) ↗

AI agents7 min text · CondensedOriginal paper ↗
Contents

Overview

At a glance

EvoSkill-GUI improves deployed interface agents by editing reusable skill files instead of model weights. An executor revises skills during a task; a separate session of the same model diagnoses failed trajectories and guides later edits. Across the evaluated backbones, maximum absolute success-rate gains are 16.2 percentage points on MobileWorld, 6.0 on AndroidWorld, and 10.5 on OSWorld. Reuse and extra inference cost are evaluated separately.

1 Introduction

Pop-ups, delayed loads, and changing widget positions make fixed plans fragile. The method stores planning, localization, recovery, and failure knowledge separately so observable mistakes can produce targeted, reusable file edits. Its claim is inference-time skill improvement without additional model training.

Related references: [24], [18], [35], [6], [11], [14], [23], [12], [39], [2], [37], [30], [33], [38], [5], [27], [29], [7], [1], [22], [34], [44], [17].

Refer to caption
Figure 1: Comparison between existing skill-based methods and EvoSkill-GUI. The upper panel illustrates the limitations of existing approaches, which rely on unstructured single-file skills, static skills during execution, external-VLM reflection, and unindexed full-text retrieval. The lower panel shows how EvoSkill-GUI addresses these issues with structured editable skill packages, metadata-based retrieval, in-rollout revision, and single-backbone self-evolution, enabling skills to be adapted and reused from real execution feedback.

2 Related Work

2.1 Agent Skills

Prior skill systems package reusable workflows, construct structured memories, or distill procedures into model weights. This framework instead revises deployed GUI skill packages from execution feedback.

Related references: [2], [43], [33], [5], [27], [38], [9], [41], [44], [30], [25], [16], [15], [37].

2.2 Self-evolving Agents

Experience-based agents update policies, memories, or reflection modules. EvoSkill-GUI specifically updates the procedural files that condition future execution.

Related references: [40], [31], [28], [22], [34], [13], [12], [29], [7], [1], [8], [41], [17].

Refer to caption
Figure 2: Overview of EvoSkill-GUI. (A) Each skill is a structured multi-file package containing accessibility utilities, planning, backup localization, recovery rules, failure cases, and metadata, accessed through a restricted tool interface. (B) The execute–reflect–revise loop alternates rollouts with in-rollout edits, isolated post-rollout critique, and skill-file revisions, with strict information isolation between executor and critic. (C) Evolved skills are indexed by metadata and reused via threshold-based scoring; misses trigger new skill generation.

3 Method

3.1 Problem Formulation

Tasks are partially observed control problems: screenshots and accessibility trees expose only part of the application state. A persistent skill conditions the agent’s policy, and revisions seek higher completion rates across related tasks.

Related references: [37].

Equations 1–3
at∼πθ​(at∣ht,S).a_{t}\sim\pi_{\theta}(a_{t}\mid h_{t},S). (1)
J⁡(S,I)=𝔼τ∼P⁡(τ∣πθ,S,ℳ,I)​[R⁡(xT)],J(S,I)=\mathbb{E}_{\tau\sim P(\tau\mid\pi_{\theta},S,\mathcal{M},I)}\left[R(x_{T})\right], (2)
S∗=arg⁡maxS​𝔼I∼𝒟​[J⁡(S,I)].S^{*}=\arg\max_{S}\;\mathbb{E}_{I\sim\mathcal{D}}[J(S,I)]. (3)

3.2 Skill Package Representation

Each package separates retrieval metadata, accessibility utilities, executable plans, backup localization, recovery rules, and failure examples. This maps a diagnosed failure to a specific component rather than rewriting one long prompt.

Equation 4
S=(D,A,P,B,C,F),S=(D,A,P,B,C,F), (4)

Tool interface.

Restricted file tools expose only the predefined skill schema. A separate read-only accessibility tool augments observations without modifying the skill.

Equation 5
𝒯={\displaystyle\mathcal{T}=\{ read,write,append,\displaystyle\texttt{read},\texttt{write},\texttt{append}, (5)
list,search,create_failure}.\displaystyle\texttt{list},\texttt{search},\texttt{create\_failure}\}.

3.3 Reflect-Revise Loop

Execution, isolated diagnosis, and targeted revision alternate until success or the iteration limit.

Rollout and instant revision.

During execution, the agent can immediately edit a skill when observed interface behavior contradicts its plan, reducing cascades from local mismatches.

Equation 6
τ(i)=Φ⁡(S(i),ℰ),\tau^{(i)}=\Phi(S^{(i)},\mathcal{E}), (6)

Reflection.

The same model acts as critic in a separate session using only the instruction and executed trajectory. It identifies failure steps, causes, and actionable fixes without access to privileged answers or the executor’s hidden reasoning.

Equations 7–8c
c(i)∼πθJ(⋅∣I,o1:T(i),a1:T(i)).c^{(i)}\sim\pi_{\theta}^{J}(\cdot\mid I,o_{1:T}^{(i)},a_{1:T}^{(i)}). (7)
𝒞J\displaystyle\mathcal{C}_{J} ⊆{I}∪o1:T∪a1:T,\displaystyle\subseteq\{I\}\cup o_{1:T}\cup a_{1:T}, (8a)
𝒞J\displaystyle\mathcal{C}_{J} ∩{S,CoTθ,GT}=∅,\displaystyle\cap\{S,\mathrm{CoT}_{\theta},\mathrm{GT}\}=\varnothing, (8b)
𝒞E\displaystyle\mathcal{C}_{E} ⊇𝒞J∪{S,CoTθ}.\displaystyle\supseteq\mathcal{C}_{J}\cup\{S,\mathrm{CoT}_{\theta}\}. (8c)

Revision.

The executor applies the critique to specific files and records failed cases. Experiments allow the initial attempt and at most two post-failure revision rounds.

Equation 9
S(i+1)∼πθ(⋅∣S(i,end),c(i),τ(i)).S^{(i+1)}\sim\pi_{\theta}(\cdot\mid S^{(i,\mathrm{end})},c^{(i)},\tau^{(i)}). (9)

3.4 Skill Retrieval and Reuse

Retrieval uses structured intent, application, platform, keywords, argument slots, history, and status. Avoiding full procedural text reduces misleading lexical overlap between unrelated tasks.

Equation 10
D=(CLOSE\displaystyle D=( id,intent,app,platform,\displaystyle\mathrm{id},\mathrm{intent},\mathrm{app},\mathrm{platform}, (10)
OPENkw,args,hist,status).\displaystyle\mathrm{kw},\mathrm{args},\mathrm{hist},\mathrm{status}).

Retrieval score.

The score combines lexical coverage and semantic tightness with application and keyword bonuses, subtracting a penalty for missing query-specific concepts. The original equations preserve the chosen weights.

Equations 11–12
base⁡(q,S)=0.6​|O||Q|+0.4​|O||Q∪TS|.\mathrm{base}(q,S)=0.6\frac{|O|}{|Q|}+0.4\frac{|O|}{|Q\cup T_{S}|}. (11)
score⁡(q,S)=clip⁡(base+bapp+bkw−pdiv,0,1).\mathrm{score}(q,S)=\mathrm{clip}(\mathrm{base}+b_{\mathrm{app}}+b_{\mathrm{kw}}-p_{\mathrm{div}},0,1). (12)

Reuse.

A candidate above the retrieval threshold is reused; otherwise a new package is generated. Packages are retained after successful execution or revision through the evolution loop.

4 Experiments

4.1 Experimental Setup

Benchmarks

Evaluation covers AndroidWorld, the GUI-only MobileWorld subset, and OSWorld desktop tasks. AndroidWorld uses 116 tasks across 20 applications; related task variants support the reuse study.

Related references: [21], [10], [32].

Models

The study includes general multimodal and GUI-specialized models, testing whether skill revision helps different model sizes and families.

Related references: [3], [20], [36], [42].

Implementations

The skill format and retrieval strategy are fixed across models. Each rollout permits 50 interaction steps, followed by at most two revision rounds after failure. Provider endpoints and locally served open models use the settings recorded in the appendix.

4.2 Main Results

Method Chrome GIMP LibreOffice Multi-Apps OS Thunderbird VLC VS Code Overall
Calc Impress Writer
General Closed-Weights Models
Claude-Sonnet-4.5 Anthropic (2026) 56.4 57.7 66.0 57.5 65.2 47.0 70.8 66.7 52.9 69.6 58.1
Qwen3-VL-Flash Bai et al. (2025) 56.4 42.3 23.0 51.0 56.5 22.0 54.2 66.7 34.4 69.6 41.6
Specialized Open-Weights Models
UI-TARS-72B-DPO Qin et al. (2025) 33.2 61.5 12.8 25.4 43.5 6.7 33.3 33.3 23.5 47.8 25.9
Computer-Use-Preview 36.9 34.6 14.9 29.7 26.1 15.8 70.8 66.7 11.8 69.6 31.2
OpenCUA-32B Wang et al. (2025a) 45.6 65.4 12.8 42.5 39.1 14.6 60.9 46.7 23.5 65.2 35.1
EvoCUA-8B Xue et al. (2026) 54.3 84.6 29.8 38.2 52.2 25.5 75.0 73.3 44.8 65.2 46.1
EvoCUA-32B 65.1 76.9 55.3 59.5 69.6 27.9 78.3 80.0 49.4 87.0 56.7
GUI-Owl-1.5-32B Ye et al. (2025b) 66.6 76.9 48.9 63.8 60.9 27.3 76.2 86.7 37.4 87.0 55.4
Skill-based Works
MMSkills Zhang et al. (2026b) 39.9 42.3 8.5 23.4 17.4 13.4 25.0 29.4 47.8 25.4
Our Method
GUI-Owl-1.5-8B 35.6 80.8 29.8 57.5 69.6 21.5 70.8 73.3 43.8 81.8 46.7
+EvoSkill-GUI 41.3 88.5 37.8 65.8 69.6 30.5 66.7 93.3 86.5 87.0 54.8
Δ\DeltaImprovement +5.7 +7.7 +8.0 +8.3 +0.0 +9.0 -4.1 +20.0 +42.7 +5.2 +8.1
Qwen3-VL-8B-Instruct 28.1 26.9 19.1 31.8 47.8 7.7 16.7 60.0 35.3 26.1 23.8
+EvoSkill-GUI 43.4 46.2 29.8 34.4 52.2 15.2 41.7 46.7 50.6 47.8 34.3
Δ\DeltaImprovement +15.3 +19.3 +10.7 +2.6 +4.4 +7.5 +25.0 -13.3 +15.3 +21.7 +10.5
Table 1: Performance comparison on the OSWorld-Verified benchmark across two models: GUI-Owl-1.5-8b and Qwen3-VL-8B-Instruct. For a fair comparison, all results are evaluated under a maximum of 50 steps.
Phase Reuse Rate Base SR Ours SR Δ\Delta
P1 (seed 30) 37.9 68.1 70.7 +2.6
P2 (seed 42) 100.0 55.2 61.2 +6.0
Table 2: Evaluation on the AndroidWorld task set. SR denotes success rate.

EvoSkill-GUI improves multiple base models on mobile GUI tasks.

On MobileWorld, Claude-Sonnet-4.6 improves from 57.1% to 67.6%, Qwen3.6-Plus from 53.3% to 69.5%, Qwen3.6-35B-A3B from 32.4% to 44.8%, and MAI-UI-8B from 29.5% to 37.1%. AndroidWorld gains are 2.6 and 6.0 percentage points for the two reported seeds.

EvoSkill-GUI extends to desktop computer-use tasks.

On OSWorld, GUI-Owl improves from 46.7% to 54.8% and Qwen3-VL-8B-Instruct from 23.8% to 34.3%. Some application-level results regress despite the overall gains.

Original figure
Figure 3: Performance comparison on MobileWorld under the GUI-only setting. The horizontal lines indicate representative reference baselines.

4.3 Ablation Study

Ablations examine skill structure, immediate revision, critic isolation, retrieval, and accessibility observations.

Skill Organization Success Rate (%)
Single-file Skill 66.67
Skill Package (Ours) 69.52
Δ\Delta Improvement +2.85
Table 3: Ablation of skill organization on MobileWorld.

Structured skill packages outperform monolithic skill files.

Separating skill components raises MobileWorld success from 66.67% for a single file to 69.52% for the structured package.

Strategy Success Rate (%)
EvoSkill-GUI (Ours) 69.52
w/o information isolation 60.95
w/o instant revision 62.86
Table 4: Ablation of information isolation and instant revision on MobileWorld on Qwen3.6-Plus.

Instant in-rollout revision prevents cascading failures.

Removing within-rollout revision lowers success from 69.52% to 62.86%, supporting correction of mismatches before a full attempt fails.

Original figure
Figure 4: Skill-library growth on the 116+116 AndroidWorld task set. EvoSkill-GUI maintains higher cumulative success rates than the baseline as the library grows and the reuse ratio increases.
(a) Component ablation of the EvoSkill-GUI skill package.
Original figure
(b) Effect of accessibility tree inputs across different models.
Original figure
Figure 5: Ablation studies on skill-package components.

Metadata-based retrieval outperforms full-text matching.

In twelve selected successful reuse cases, metadata retrieval finds the intended skill in all twelve at threshold 0.6; full-text matching finds one. This small selected-case analysis is supplemented by held-out retrieval evaluation.

Information isolation improves reflection quality.

Removing critic information isolation reduces success from 69.52% to 60.95%. Separate-session diagnosis helps avoid inheriting the executor’s mistaken assumptions.

Original figure
Figure 6: Three round evolution of multiple base models on MobileWorld.
Task name Metadata Score Full-text Score Meta@0.6 Full-text@0.6
Mastodon New Post 0.8036 0.3431 ✓ ✗
Mattermost Create Channel 1.0000 0.4769 ✓ ✗
TaoDian Item Checkout 1.0000 0.6182 ✓ ✓
Mastodon Reply (Moussaka) 1.0000 0.3833 ✓ ✗
Change Wallpaper (Sunflower) 0.8333 0.4184 ✓ ✗
Mattermost Send File 0.7003 0.3414 ✓ ✗
Mattermost Reply Result 0.7833 0.3521 ✓ ✗
Check Cart Price 0.9500 0.4231 ✓ ✗
Mastodon Unfollow 0.8167 0.3876 ✓ ✗
Mastodon Save Photos 0.8500 0.3654 ✓ ✗
Google Maps Alibaba South 0.9167 0.4512 ✓ ✗
Mastodon Conditional Favorite 0.8500 0.2987 ✓ ✗
Average / Hit Rate 0.8753 0.4050 12/12 1/12
Table 5: Ablation of metadata for skill retrieval on MobileWorld reuse cases, all cases are completed successfully by EvoSkill-GUI. @0.6 indicate whether the corresponding retrieval score exceeds the threshold θr=0.6\theta_{r}=0.6.

Each skill-package component contributes to performance.

Removing planning documents reduces success to 65.71%; removing failure reflection reduces it to 66.67%. The components provide complementary guidance.

Accessibility-tree utilities benefit grounding across models.

Accessibility-tree inputs improve success by 1.90–3.80 percentage points across the compared backbones, complementing visual grounding.

A moderate retrieval threshold balances skill reuse and selectivity.

Threshold 0.6 performs best in the reported sweep. A lower threshold admits weak matches; a higher threshold rejects useful experience.

Retrieval Threshold θr\theta_{r} Success Rate (%)
0.4 55.2
0.6 (Ours) 69.5
0.8 66.7
Table 6: Sensitivity of MobileWorld performance to the retrieval threshold θr\theta_{r}.

4.4 Analysis

Skill evolution provides cumulative gains and largely saturates after three rounds.

Most improvement arrives within three execution rounds. In the extended Qwen3.6-Plus study, the first three rounds gain 12.4 points and two additional rounds add only 0.9.

Skill Source Tasks Recovered Recovery (%)
New skills 72 8/24 33.3
Reused skills 160 19/65 29.2
Overall 232 27/89 30.3
Table 7: Analysis of skill-reuse on failure-to-success recovery on the AndroidWorld.
Retrieval Threshold θr\theta_{r} Success Rate (%)
0.4 55.2
0.6 (Ours) 69.5
0.8 66.7
Table 8: Sensitivity of MobileWorld performance to the retrieval threshold θr\theta_{r}.
Failure Type Target File Recovered Rate
Plan-level error plan.md 16/23 69.6
Grounding error backup.md 6/12 50.0
Missing contingency recover.md 1/4 25.0
Table 9: Repairability analysis of failed executions by target skill component.

Skill evolution outperforms repeated sampling under a comparable budget.

Under a comparable token budget, three evolution rounds reach 69.5% versus 62.8% for three independent baseline attempts, using 94.75 million versus 103.00 million tokens.

Evolved skill libraries remain reusable and effective on related tasks.

A library built on 116 AndroidWorld tasks benefits 116 related variants: success rises from 68.1% to 70.7% in phase one and from 55.2% to 61.2% in phase two. These results concern related parameterized tasks rather than arbitrary new applications.

Skill libraries accumulate reusable experience over task streams.

The sequential task stream accumulates more reusable packages, with the reported library reaching 98 skills. Reuse statistics depend on whether the paper measures a phase, stream, or selected subset.

Different failure types require different skill edits.

Plan edits repair 16 of 23 analyzed planning failures; localization edits repair 6 of 12 grounding failures, and recovery edits repair 1 of 4 missing-contingency failures. Repairability differs by error type.

5 Conclusion

Structured, editable skills can improve GUI execution without changing model weights. The gains come from targeted revision and reuse under the tested inference budgets.

Limitations

Incorrect perception or diagnosis can produce harmful edits, and no formal verifier currently rejects regressions. Benchmark tasks do not cover all interface changes or personal configurations. Failed attempts incur extra critic and revision calls; benefits depend on later reuse.

Ethics Statement

The authors discuss authorization, sandboxing, auditability, and review of reusable skills for deployment beyond benchmarks. They report no collection of personal user data in these experiments.

Appendix A Appendix

A.1 Complete Implementation Details

The appendix records provider-based and local serving configurations, shared evaluation settings, and research-use conditions.

Model serving.

The reported local configuration uses vLLM, bfloat16, a 65,536-token context, and at most five images per prompt. Exact serving parameters remain in the original implementation description.

Evaluation.

Runs permit 50 steps with a three-second per-step wait, and up to three execution iterations including the initial attempt.

Environment.

Environment settings include parallel benchmark containers, model-specific screenshot resizing, and normalized action coordinates. These choices affect runtime and reproducibility.

Excluded tasks.

Twelve MobileWorld Google-proxy tasks are excluded for evaluation stability.

A.2 Held-out skill transfer and retrieval quality

Evolved skills transfer to held-out tasks.

A library constructed on 87 MobileWorld tasks is tested on 30 held-out tasks. EvoSkill-GUI solves 22/30 versus 12/30 for pass@3. Five additional successes use retrieved skills and five use newly generated packages.

Retrieval result Tasks Ours S/F pass@3 S/F
Retrieved existing skill 16 12/4 7/9
No retrieval 14 10/4 5/9
Total 30 22/8 12/18
Table 10: Held-out MobileWorld results. S/F denotes successful/failed tasks.

Metadata-based retrieval remains precise on held-out tasks.

Of sixteen held-out tasks triggering retrieval, fifteen receive an appropriate skill. The remaining mismatch illustrates possible negative transfer.

A.3 Same-backbone self-diagnosis

Same-backbone critique preserves self-contained evolution.

Using one backbone for execution and critique avoids attributing improvement to a stronger supervisor, but critique still consumes inference tokens.

Information isolation separates critique from execution.

Critic sessions see the instruction and observed trajectory, excluding the skill package, executor CoT, and ground truth.

Same-backbone diagnosis is meaningful but imperfect.

Manual review finds 115 correct judgments among 143, or 80.4%. At least four false success judgments prematurely stop revision, demonstrating a concrete weakness of self-diagnosis.

Judgment Count Rate (%)
Correct diagnosis 115/143 80.4
Incorrect diagnosis 28/143 19.6
Table 11: Manual audit of same-backbone critic judgments across 56 failed tasks.

A.4 Analysis of negative OSWorld results

Two OSWorld application/model entries show negative changes despite overall improvement.

Rare negative cases are sensitive to run-level variation.

Only three tasks account for the negative changes, and reruns vary. Small application-level differences therefore provide weak evidence about systematic regressions.

Skill guidance can overcomplicate direct operations.

For short direct tasks, an imperfect skill can add unnecessary steps or reinforce an incorrect procedure. More guidance is not always helpful.

A.5 Token cost under repeated sampling.

The corrected accounting reports 94.75 million tokens for EvoSkill-GUI versus 103.00 million for pass@3. An earlier 168-million figure double-counted intermediate cumulative totals.

Category Parameter Value
Execution Temperature 0.0
Execution Max output tokens 8192
Execution History screenshots 3
Execution Coordinate scale factor 1000
Execution Max skill-tool calls/ step 4
Skill generation Temperature 0.3
Skill generation Max output tokens 4096
Critic Temperature 0.0
Critic Max output tokens 4096
Critic Max screenshots 8
Critic Screenshot max dimension 1280
Retrieval BM25 k1k_{1} 1.5
Retrieval BM25 bb 0.75
Retrieval Keyword threshold 0.6
Retrieval Divergence threshold 0.65
Retrieval Divergence weight 0.3
Retrieval Domain bonus 0.15
Retrieval Keyword bonus ≤\leq 0.2
Revision Revise order P →\rightarrow B →\rightarrow C
Revision Failure examples 3 at most
Table 12: Main implementation hyperparameters used in EvoSkill-GUI.
Method Total (M) Avg./ Task (M) Avg./ Round (M) Acc. (%)
Baseline pass@1 33.20 0.284 0.284 53.3
Baseline pass@3 103.00 0.880 0.293 62.8
Single-file skill pass@3 84.54 0.723 0.372 66.7
EvoSkill-GUI (3 rounds) 94.75 0.810 0.439 69.5
Table 13: Test-time token cost and accuracy on MobileWorld under pass@1 and pass@3 evaluation. M denotes million tokens.

Accessibility trees add moderate runtime overhead but little token usage overhead.

Accessibility trees add collection and processing time while token usage remains broadly similar. Their cost is mainly runtime overhead in the reported examples.

Task Token cost(K) Time(s)
AcceptMeetingTask 67.2 / 71.1 437 / 288
MastodonNewPostTask 79.8 / 79.4 476 / 358
SetAlarmTask 189.1 / 186.0 526 / 405
ScheduleLunchViaSmsTask 800.0 / 820.1 1277 / 1125
Table 14: Comparison of token usage and execution time with / without accessibility-tree inputs.

A.6 Case Studies

Two case studies trace an observed failure through critic diagnosis, file edits, and later execution.

A.6.1 Case 1: Verification Checkpoint Insertion

In an email task, locating a resume file was mistaken for verifying and attaching it. The revised plan adds explicit document and attachment checks, and the next rollout succeeds.

Refer to caption
(a) Failed rollout.
Refer to caption
(b) Successful rollout by EvoSkill-GUI.
Figure 7: Case study on SendInterviewEmailTask. EvoSkill-GUI revises the skill from a locate-only workflow into a locate–verify–attach workflow.
Round Score Steps Failure Mode
1 0.0 14 Skipped file verification
2 1.0 14 Success
Table 15: Evolution summary for SendInterviewEmailTask.

A.6.2 Case 2: Completeness Enforcement for List Traversal

In a cart task, incomplete scrolling produces the wrong top-three price total. A revised traversal plan alone fails again; recording the concrete failed shortcut and enforcing item-count completion leads to success on the third attempt.

Refer to caption
(a) Failed rollout.
Refer to caption
(b) Successful rollout by EvoSkill-GUI.
Figure 8: Case study on CheckCartPriceTask. EvoSkill-GUI learns to verify complete traversal before reporting the top-3 item total.
Round Score Steps Answer GT
1 0.0 25 2657 13186
2 0.0 28 2657 13186
3 1.0 11 13186 13186
Table 16: Evolution summary for CheckCartPriceTask.

A.6.3 Summary

The examples show how failure evidence changes specific planning, localization, and recovery files instead of merely prompting an independent retry.

A.6.4 Visualized Diagnosis-to-Edit Paths

Case 1: SendInterviewEmailTask

Round 1 Failure Trace • Finds Kevin_CV.pdf in Downloads. • Leaves Files without opening the PDF. • Sends email without content verification or attachment. Critic Diagnosis Failure type: missing verification checkpoint.
Root cause: the agent treats “filename located” as “resume verified”.
Required correction: insert open–verify–attach checkpoints before email sending.
Round 2 Effect Locate-only workflow →\rightarrow locate–verify–attach workflow. The revised skill succeeds in 14 steps.

Case 2: CheckCartPriceTask

Round 1 Failure Trace • Scans only part of a 20-item cart. • Computes top-3 total from visible items. • Outputs 2657 instead of 13186. Critic Diagnosis after Round 1 Failure type: incomplete traversal.
Root cause: scrolling is performed, but completeness is never verified.
Required correction: count unique items and stop only after end-of-list or total count is confirmed.
Round 2 Failure Trace • Repeats answer 2657. • Scrolls extensively but does not count 20 unique items. • Fails to prove that no unseen items remain. Critic Diagnosis after Round 2 Repeated root cause: premature answer before complete traversal.
Required correction: preserve the failure as an explicit negative example.
Round 3 Effect Partial scan shortcut →\rightarrow systematic traversal with failure memory. The revised skill reports 13186 in 11 steps.

A.7 Algorithm Pseudocode

Algorithm 1 EvoSkill-GUI
0:  Instruction II, Environment ℰ\mathcal{E}, Skill Library ℒ\mathcal{L}, Policy πθ\pi_{\theta}, Critic πθJ\pi_{\theta}^{J}, Max iterations NN, Threshold θr\theta_{r}
0:  Task completion trajectory τ\tau, Updated Skill Library ℒ\mathcal{L}
1:  Compute score​(I,S)\text{score}(I,S) for all S∈ℒS\in\mathcal{L} ⊳\triangleright Stage 1: Retrieval (Eq. 12)
2:  if max⁡score​(I,S)>θr\max\text{score}(I,S)>\theta_{r} then
3:   S(0)←RetrieveTop​(I,ℒ)S^{(0)}\leftarrow\text{RetrieveTop}(I,\mathcal{L}) ⊳\triangleright Reuse verified skill package
4:  else
5:   S(0)←InitSkill​(D,A,P,B,C,F)S^{(0)}\leftarrow\text{InitSkill}(D,A,P,B,C,F) ⊳\triangleright Create new skill package (Eq. 4)
6:  end if
7:  ⊳\triangleright Stage 2: Self-evolution loop
8:  for i=0i=0 to N−1N-1 do
9:   τ(i),S(i,end)←Φ⁡(S(i),ℰ)\tau^{(i)},S^{(i,\mathrm{end})}\leftarrow\Phi(S^{(i)},\mathcal{E}) ⊳\triangleright Rollout with instant revision (Eq. 6)
10:   if τ(i)\tau^{(i)} succeeds then
11:    ℒ←RegisterOrUpdate​(ℒ,S(i,end))\mathcal{L}\leftarrow\text{RegisterOrUpdate}(\mathcal{L},S^{(i,\mathrm{end})}) ⊳\triangleright Register verified skill
12:    return τ(i),ℒ\tau^{(i)},\mathcal{L}
13:   else
14:    c(i)∼πθJ(⋅∣I,o1:T(i),a1:T(i))c^{(i)}\sim\pi_{\theta}^{J}(\cdot\mid I,o_{1:T}^{(i)},a_{1:T}^{(i)}) ⊳\triangleright Reflect: isolated diagnosis (Eq. 7)
15:    S(i+1)∼πθ(⋅∣S(i,end),c(i),τ(i))S^{(i+1)}\sim\pi_{\theta}(\cdot\mid S^{(i,\mathrm{end})},c^{(i)},\tau^{(i)}) ⊳\triangleright Revise: edit skill package (Eq. 9)
16:    S(i+1).F←S(i+1).F∪{(I,c(i),τ(i))}S^{(i+1)}.F\leftarrow S^{(i+1)}.F\cup\{(I,c^{(i)},\tau^{(i)})\} ⊳\triangleright Append failure case via create_failure
17:   end if
18:  end for
19:  ℒ←RegisterOrUpdate​(ℒ,S(N))\mathcal{L}\leftarrow\text{RegisterOrUpdate}(\mathcal{L},S^{(N)}) ⊳\triangleright Save revised skill package
20:  return τ(N−1),ℒ\tau^{(N-1)},\mathcal{L}

A.8 Prompt Template

A.8.1 Executor System Prompts

EXECUTOR_SKILL_GENERATOR_SYSTEM_PROMPT_JSON EXECUTOR_SKILL_GENERATOR_SYSTEM_PROMPT_XML

A.8.2 Skill Generation Prompts

SEQUENTIAL_SKILL_GEN_STEP_1_META_PLAN_PROMPT SEQUENTIAL_SKILL_GEN_STEP_2_BACKUP_PROMPT SEQUENTIAL_SKILL_GEN_STEP_3_RECOVER_PROMPT

A.8.3 Critic Reflection Prompts

CRITIC_REFLECTION_SYSTEM_PROMPT CRITIC_REFLECTION_USER_PROMPT

A.8.4 Skill Revision Prompts

SKILL_REVISION_SYSTEM_PROMPT
Concept demonstration

Open image ↗