| --- |
| license: apache-2.0 |
| --- |
| # Functional Semantics Embedding of GUI Screens for Knowledge-Augmented GUI Agents |
|
|
| This dataset is a **benchmark** used in the experiments of the paper "**Functional Semantics Embedding of GUI Screens for Knowledge-Augmented GUI Agents**," submitted to **ACM MM 2026**. |
|
|
| This benchmark also includes MLLM prompts for step-level action decision, as well as prompts used to generate knowledge. |
|
|
| --- |
|
|
| ## Download |
|
|
| ```bash |
| hf download user83kd9x/knowledge_agent_benchmark \ |
| --repo-type dataset \ |
| --local-dir knowledge_agent_benchmark |
| ``` |
|
|
| --- |
|
|
| ## data.zip |
|
|
| All data are organized at the task episode level. Under each directory named after an episode ID, the data for each step are stored as individual JSON files. |
|
|
| ### Data Structure |
|
|
| Each JSON file contains the following data. |
| - step_id |
| - episode_id |
| - app_name |
| - task_objective: The overall objective of the task trajectory. |
| - expected_result: The screen expected upon successful |
| completion of the task. |
| - previous_action_history: A sequence of sentences sum- |
| marizing the actions taken at all preceding steps. |
| |
| - actions: Ground truth actions to be performed at the current step. If multiple actions can be considered correct, up to two are provided as valid answers. |
| - action_type |
| - box: The bounding box of the target GUI element or a specific region associated with the action. |
| - direction: The direction of the action, in the case of a Scroll or Swipe. |
| - text: For an Input action, the text that should be entered. |
| - wait_duration: For a Wait action, the duration to wait (in seconds). |
| |
| - knowledges: Knowledge associated with each ground truth action. |
| - box: The bounding box of the target GUI element or a specific region associated with the action. |
| - appearance: A visual description of the target GUI element or region. |
| - function: A functional description inferred from the preand post-action screenshots. |
| |
| --- |
| |
| ## screenshots.zip |
| |
| All screenshot images for every step are stored in this archive. The images are also organized by episode, and each image file is named using its corresponding step ID. |
| |
| --- |
| |
| ## Test Prompt |
| |
| `prompts/test_without_knowledge.py` and `prompts/test_with_knowledge.py` are prompts for performing step-level action decision using the given benchmark data and a single screenshot. |
| |
| The only difference between the two prompts is whether knowledge is included; all other aspects are identical. |
| |
| Both prompts include the output of the merge detect API from the repository below to assist the MLLM’s coordinate grounding. For details on how to use the merge detect API, please refer to the repository below. |
| |
| https://github.com/conference-anonymous-author/Screen-SBERT.git |
| |
| The knowledges included in the `PROMPT_TEST_WITH_KNOWLEDGE` prompt is not the step-specific knowledge stored in the JSON files. Instead, it consists of the merged knowledge set obtained by embedding the current screenshot and retrieving the nearest cluster based on similarity. |
|
|
| --- |
|
|
| ## Knowledge Generation Prompt |
|
|
| The prompts in `prompts/knowledge_prompt.py` are used to generate knowledge from the benchmark data. All knowledge data included in the JSON files were generated using these prompts. |
|
|
| For Click, Scroll, and Swipe actions, if a result screenshot is available, both the before- and after-action screenshots are provided to generate knowledge. In this case, the current screenshot is annotated with a bounding box around the target element or region of the action. Further details are described in the prompt. |
|
|
| In this benchmark, some steps contain more than one action because there may exist a more efficient action than the one that leads to the next screenshot. In such cases, both actions are treated as correct answers. Since no result screenshot exists for the more efficient action, a single screenshot is provided along with the `PROMPT_EFFICIENT_ACTION_KNOWLEDGE` prompt. The `{best_action}` field contains the action type and its associated parameters. |
