user83kd9x/screenshot_clustering

Functional Semantics Embedding of GUI Screens for Knowledge-Augmented GUI Agents

This repository contains the implementation code for the clustering algorithm and knowledge aggregation proposed in the paper "Functional Semantics Embedding of GUI Screens for Knowledge-Augmented GUI Agents," submitted to ACM MM 2026.

---

Download

hf download user83kd9x/screenshot_clustering \
  --repo-type dataset \
  --local-dir screenshot_clustering

---

embeddings.zip

All screenshots from the benchmark used in this study have been embedded using the Screen-SBERT model.

The embedding files are stored in NumPy format and are organized by application. Each file name corresponds to a step ID in the benchmark.

---

cluster_embeddings.py

You can perform clustering directly as shown below. (A NumPy library is required in your Python environment.)

python cluster_embeddings.py \
    --embed-dir embeddings \
    --threshold 0.6210

As a result, the embedding vectors are clustered for each application, and text files are generated containing the list of step IDs belonging to each cluster, along with the representative ID closest to the cluster center.

clustering.zip

In addition to the outputs generated by the code above, we also include the screenshot images to enable visual inspection of the clustering results.

---

knowledges.zip

The knowledge associated with each step ID in the benchmark below is organized according to the clustering results.

https://huggingface.co/datasets/user83kd9x/knowledge_agent_benchmark

Within this directory, the function field of the knowledge instance is stored under the key "Knowledge".

merge_clustered_knowledges.py

python3 merge_clustered_knowledges.py \
    --knowledges-dir knowledges

This code removes knowledge entries with overlapping bounding box regions within each cluster to generate the final set of knowledge used by the agent system.

In all knowledge sets, bounding box data is converted into center coordinate representations (to reduce the number of tokens during prompt augmentation).