Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents

The authors propose EvoSkill-GUI, a training-free framework that enables GUI agents to dynamically update and refine their procedural skills through a reflect-revise-reuse loop during deployment. Each skill is structured as a multi-file package containing retrieval metadata, execution plans, and failure-recovery rules. An executor performs instant in-rollout revisions while an isolated critic diagnoses failures. Experiments across MobileWorld, AndroidWorld, and OSWorld benchmarks demonstrate consistent performance improvements over base models without any additional training.

EvoSkill-GUI achieves maximum gains of +16.2%, +6.0%, and +10.5% across MobileWorld, AndroidWorld, and OSWorld benchmarks respectively.

The framework operates entirely without training by leveraging direct skill-file editing through a restricted tool interface.

Evolved skill libraries continue to benefit related tasks rather than requiring reconstruction from scratch.