Abstract
Generative Synthesis of Web Environments with Controllable Difficulty and Integrated Validation for Agent Training
ABSTRACT
Training and evaluating computer control agents may be challenging due to limitations in the availability of large-scale, diverse, and controllable digital environments. Existing methods can have limitations related to scalability, reproducibility, and the ability to parameterize task difficulty. A described framework can utilize generative models, such as large language models, to synthesize functional, web-based environments. The framework can generate source code for an environment (e.g., HTML, CSS, JavaScript), a corresponding natural language task description, and a validation function to programmatically assess task completion. This approach can provide a scalable and reproducible method for creating varied datasets. This capability can support features such as granular control over environmental and task difficulty, which may be used for training strategies, for example, curriculum learning, to aid in the development of computer control agents.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Riva, Oriana and Li, Alice, "Generative Synthesis of Web Environments with Controllable Difficulty and Integrated Validation for Agent Training", Technical Disclosure Commons, (March 09, 2026)
https://www.tdcommons.org/dpubs_series/9475