saelyne/GuideBench

GUIDE: A Benchmark for Understanding and Assisting Users in Open-Ended GUI Tasks

GUIDE (GUI User Intent Detection Evaluation) is a benchmark that evaluates multimodal models on their ability to perceive user behavior, infer intent, and provide assistance in open-ended GUI tasks. It consists of 67.5 hours of screen recordings from 120 novice user demonstrations with think-aloud narrations, across 10 software applications.

Saelyne Yang, Jaesang Yu, Yi-Hao Peng, Kevin Qinghong Lin, Jae Won Cho, Yale Song, Juho Kim. GUIDE: A Benchmark for Understanding and Assisting Users in Open-Ended GUI Tasks. CVPR 2026.

---

Benchmark Tasks

GUIDE defines three evaluation tasks that progress from perception to reasoning to assistance:

#	Task	Description	Format
1	User Behavior State Detection	Identify the user's current behavioral state from a video segment	9-way classification
2	Intent Prediction	Infer the user's immediate goal within the segment	4-way MCQ
3	Help Prediction	Determine whether and what kind of help the user needs	Binary + 4-way MCQ

Dataset Structure

GuideBench/
├── annotations/
│   ├── 1_taxonomy_annotations.json
│   ├── 2_intent_annotations.json
│   └── 3_help_annotations.json
├── tasks/
│   ├── Canva_A_Task_1.md
│   ├── Canva_A_Task_2.md
│   └── ...
├── transcriptions/
│   ├── transcripts_Canva_A_Task_1_8_1.json
│   └── ...
└── videos/
    ├── video_urls.csv
    └── segment_urls.csv

---

Annotations (annotations/)

Each entry corresponds to a video segment and annotations for the benchmark tasks.

1. User Behavior State Detection (1_taxonomy_annotations.json)

Field	Description
video_id	{Software}_{Group}_Task_{N}_{FileID} (e.g., Premiere_Pro_A_Task_1_7_1)
start_time	Segment start time in seconds
end_time	Segment end time in seconds
label	One of the 9 behavior states from the taxonomy

2. Intent Prediction (2_intent_annotations.json)

Field	Description
video_id	{Software}_{Group}_Task_{N}_{FileID} (e.g., Premiere_Pro_A_Task_1_7_1)
start_time	Segment start time in seconds
end_time	Segment end time in seconds
options	Dict of 4 answer choices (A–D)
answer	Ground-truth answer key (A–D)

3. Help Prediction (3_help_annotations.json)

Field	Description
video_id	{Software}_{Group}_Task_{N}_{FileID} (e.g., Premiere_Pro_A_Task_1_7_1)
start_time	Segment start time in seconds
end_time	Segment end time in seconds
options	Dict of 4 answer choices (A–D) for help content
answer	Ground-truth answer key (A–D)
type	One of three help types (see below)

Help types:

Type	Description
1_explicit	Explicit help-seeking behavior (e.g., switching to Google, YouTube, or ChatGPT)
2_implicit	Implicit help-seeking signals inferred from user narration (e.g., expressing confusion or uncertainty)
3_no_help	User is confidently working without needing assistance, as indicated by narration

---

Tasks (tasks/)

Markdown files describing the open-ended tasks given to participants. Each file corresponds to one task condition.

File naming: {Software}_{Group}_Task_{N}.md (e.g., Canva_A_Task_1.md, Premiere_Pro_B_Task_2.md) Each application has 4 open-ended tasks (Groups A/B × Tasks 1/2), completed by 3 different novice users each (120 demonstrations total).

---

Transcriptions (transcriptions/)

Think-aloud narrations of screen recordings transcribed using WhisperX.

File naming: transcripts_{Software}_{Group}_Task_{N}_{FileID}.json (e.g., transcripts_Canva_A_Task_1_8_1.json)

---

Videos (videos/)

It contains two CSV files with URLs to the screen recordings hosted on S3:

video_urls.csv

Full-length, privacy-redacted screen recordings.

Column	Description
video_id	{Software}_{Group}_Task_{N}_{FileID} (e.g., Premiere_Pro_A_Task_1_7_1)
file_name	S3 URL to the full video (.mp4)

segment_urls.csv

Privacy-redacted annotated segments for the remaining recordings.

Column	Description
video_id	{Software}_{Group}_{Task_#}_{File_ID} (e.g., Premiere_Pro_A_Task_1_7_1)
file_name	S3 URL to a redacted annotated segment (.mp4)

Segment file names follow the format {video_id}_segment_{start_time}_{end_time}.mp4 (e.g., Canva_A_Task_1_8_1_segment_1409.515_1433.191.mp4).

Release status: We initially release 50 full videos and annotated segments for the remaining 70 videos. We will incrementally release additional full videos as they pass our privacy review process (PII screening and redaction). Check back for updates.

---

Privacy & Personally Identifiable Information (PII) Reporting

All released videos have undergone PII detection and redaction. However, if you encounter any personally identifiable information (phone number, email address, face, etc.) in the recordings, please report it using the form below so we can address it promptly.

🔒 Report PII in GUIDE recordings

---

Citation

If you use GUIDE in your research, please cite:

@inproceedings{yang2026guide,
  title     = {GUIDE: A Benchmark for Understanding and Assisting Users in Open-Ended GUI Tasks},
  author    = {Yang, Saelyne and Yu, Jaesang and Peng, Yi-Hao and Lin, Kevin Qinghong and Cho, Jae Won and Song, Yale and Kim, Juho},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.