The dataset viewer should be available soon. Please retry later.
EVIDENT: Task Metadata for Evidence-Grounded GUI Evaluation
This release contains 220 GUI/web agent task records for evidence-grounded trajectory evaluation: 150 Korean and 70 English tasks. Task definitions pair a user request with its intended outcome and evidence checks.
Evaluation code and instructions: yerinNam/EVIDENT.
arXiv Paper: arxiv_EVIDENT.
βοΈ Built by Human Annotators
π οΈ Task and evidence metadata construction
All 220 tasks and their evidence metadata were written by two annotators. Each task was drafted by one annotator and cross-checked by the other, and disagreements were resolved through discussion. Construction took about two months (approximately 4 hours/day, 5 days/week per annotator; roughly 320 annotator-hours in total).
π Metadata validation
Three additional annotators who were not involved in construction validated 50 randomly sampled tasks. Two of them independently re-wrote the evidence checks without seeing the original metadata, and the third compared the two sets for semantic equivalence and audited the original metadata against the annotation guidelines.
| Check | Result |
|---|---|
| Independent re-authoring agreement | 48/50 (96%) |
| Metadata audit (observability, necessity, answer-independence) | 49/50 (98%) |
βοΈ Judge validation
Three annotators independently relabeled 126 stratified agent trajectories (both languages, all three task types) with the four outcome labels: Task Failure, Hallucinated Success, Claim-Ungrounded Success, and True Success. The main judge (Qwen3-VL-235B) agrees with the human labels on 91.27% of trajectories (Cohen's ΞΊ = 0.864).
Validation and relabeling took about two days in total.
π₯ Who are the annotators?
The annotators are native Korean speakers proficient in English.
Two annotators constructed the tasks and evidence metadata. Three additional annotators who were not involved in construction performed the metadata validation. Judge validation involved three annotators who independently relabeled the agent trajectories.
π Dataset statistics
Task-type distribution
| Task type | Korean | English | Total |
|---|---|---|---|
| Semantic Grounding | 50 | 27 | 77 |
| Contextual Inference | 50 | 21 | 71 |
| Visual Grounding | 50 | 22 | 72 |
| Total | 150 | 70 | 220 |
π Results
| Model | Backbone | EN TSR | EN ESR | EN GSR | KO TSR | KO ESR | KO GSR |
|---|---|---|---|---|---|---|---|
| Gemini CUA | gemini-3-flash-preview | 87.14 | 70.00 | 55.71 | 81.33 | 70.67 | 45.33 |
| Claude CUA | claude-opus-4-7 | 80.00 | 57.14 | 35.71 | 74.00 | 42.00 | 30.00 |
| GPT CUA | gpt-5.4 | 70.00 | 61.43 | 51.43 | 56.67 | 44.00 | 27.33 |
| MolmoWeb-8B-Native | allenai/MolmoWeb-8B-Native | 57.14 | 51.43 | 45.71 | 16.00 | 10.00 | 4.67 |
| Fara-7B | microsoft/Fara-7B | 57.14 | 50.00 | 35.71 | 8.00 | 0.67 | 0.00 |
| UI-TARS-1.5-7B | ByteDance-Seed/UI-TARS-1.5-7B | 54.29 | 35.71 | 22.86 | 24.00 | 13.33 | 10.00 |
- TSR (Task Success Rate): the final answer completes the task.
- ESR (Evidence Success Rate): task success + the required evidence was observed during navigation.
- GSR (Grounded Success Rate): evidence success + the final answer's core claims are supported by the observed evidence.
By construction, GSR β€ ESR β€ TSR. All agents were run in the same MolmoWeb-based browser pipeline with a 50-step cap (May 2026) and scored by Qwen3-VL-235B as the judge. Per-task-type results are reported in the paper.
Contents
tasks/150_ko_metadata.json: 150 Korean task records.tasks/70_en_metadata.json: 70 English task records.
Both files are JSON arrays. Every record includes id, prompt, task_type,
start_url, user_intent, and evidence_checks (a list of checks).
Some records also include visual_target (a string) or visual_cues (a list).
IDs are unique within each language file. These are task metadata files;
agent trajectories and screenshots must be collected separately.
Download
from huggingface_hub import snapshot_download
snapshot_download("KORMo-VL/EVIDENT", repo_type="dataset", local_dir="EVIDENT-data")
Evaluate agent trajectories
Use the EVIDENT evaluation code to evaluate agent trajectories against the released task metadata. See the repository README for setup and evaluation instructions.
Contact
- KyungTae Lim, Professor at KAIST.
ktlim@kaist.ac.kr
Contributor
- Yerin Nam (yerinnam@kaist.ac.kr)
- Minkyung Cho (minkyung.cho@kaist.ac.kr)
- Minkyu Kim (minkyukim@kaist.ac.kr)
- Seungwoo Song (sswoo@kaist.ac.kr)
- Hangyeol Yoo (hg.yoo@kaist.ac.kr)
- Gyun Lee (gyunini@kaist.ac.kr)
- KyungTae Lim (ktlim@kaist.ac.kr)
(admin@univa.ai) Β· Website
Acknowledgments
- This research was supported by the Open Source AI/SW Development and Utilization Support Project of the Ministry of Science and ICT (MSIT) and the National IT Industry Promotion Agency (NIPA), Republic of Korea (Grant No. PJT-26-040055).
- This work utilized GPU resources provided through the Advanced GPU Utilization Support Program (KyungTae Lim).
Citation
@inproceedings{nam2027evident,
title = {{EVIDENT}: Evaluating Evidence Grounding in {GUI} Agents},
author = {Nam, Yerin and Cho, Minkyung and Kim, MinKyu and
Song, SeungWoo and Yoo, Hangyeol and Kim, Jihyo and
Lim, KyungTae},
booktitle = {Findings of the Association for Computational Linguistics: EACL 2027},
year = {2027},
publisher = {Association for Computational Linguistics},
note = {Accepted for publication; in press},
url = {https://openreview.net/forum?id=PqIPwEDcJw}
}
- Downloads last month
- 146