Elamur Research⌕
☰
home / papers / 2026 / Seahawk at MedGenVidQA 2026: LLM Segment-Range Selection for Medical Visual Answer Localization

Seahawk at MedGenVidQA 2026: LLM Segment-Range Selection for Medical Visual Answer Localization

Paper · 2026 · Xiaotian Tian and Gulustan Dogan

Summary

This paper describes the Seahawk team's entry for MedGenVidQA 2026 Task C, which focuses on locating answers within videos. For each pairing of a medicine-related question with an instructional video, the system must find the time interval in which the answer is shown or explained. The system works only on text. It splits the timestamped ASR transcript into numbered segments, and DeepSeek picks a contiguous range by returning just the start and end segment IDs. It never writes timestamps itself. The final times come from the boundaries of the chosen segments, which prevents invented or malformed outputs. For long transcripts, the system queries overlapping windows and scores each candidate range by lexical overlap with the question minus a duration penalty, which favors short spans. On the 80-instance test set, it reached about 52.3% mIoU with no training or fine-tuning. The work shows that a constrained, prompt-only LLM pipeline is a cheap and traceable baseline for temporal grounding in medical videos. Its main weaknesses are ASR errors, the lack of visual input, and answers that cross window boundaries.

Key points

  • The LLM returns only START_SEGMENT_ID and END_SEGMENT_ID. Start times are floored and end times are ceiled to MM:SS, and results are written to task_c_submission.json.
  • A sanity check on the first 20 items of MedVidQA/test.json compared two setups. A setup favoring complete, expanded ranges scored mIoU 0.3217. Short spans with a duration penalty scored 0.4815 (length_penalty = 0.06), and a similar short-span variant scored 0.4689.
  • Full Task C test results (C1–C80) were: mIoU 52.31%, IoU@0.3 71.25%, IoU@0.5 52.50%, and IoU@0.7 42.50%.
  • In the shared-task comparison of six systems, Seahawk (listed as TXT66) ranks fourth by IoU@0.7, behind LAMAR-2 (mean IoU 77.50) and NJUST-KMG (75.48).
  • The authors list several limitations: hyperparameters were tuned on only 20 samples, the method assumes contiguous answer spans, the duration penalty sometimes cuts off answers that are genuinely long, and null outputs are filled with placeholders.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Nankai UniversityTianjin · China24 works · 762 citations

Also on the byline: University of North Carolina Wilmington.

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksInformation Systems · 1,437 works

Cite

Cite as Xiaotian Tian and Gulustan Dogan. 2026. Seahawk at MedGenVidQA 2026: LLM Segment-Range Selection for Medical Visual Answer Localization. Proceedings of the BioNLP 2026 (Shared Tasks).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bionlp-2.34

@inproceedings{tian2026seahawk,
  title = {{Seahawk at MedGenVidQA 2026: LLM Segment-Range Selection for Medical Visual Answer Localization}},
  author = {Xiaotian Tian and Gulustan Dogan},
  booktitle = {Proceedings of the BioNLP 2026 (Shared Tasks)},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-2.34},
  url = {https://doi.org/10.18653/v1/2026.bionlp-2.34}
}

References

4 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Deepak Gupta, Kush Attal, and Dina Demner-Fushman. 2023. A dataset for medical instructional video classification and question answering. Scientific Data, 10(1):158. find
  2. Deepak Gupta, Kush Attal, and Dina Demner-Fushman. 2024. Towards answering health-related questions from medical videos: Datasets and approaches. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 16399–… find
  3. Deepak Gupta, Collin Scott Campbell, Pedram Golnari, and Dina Demner-Fushman. 2026. Overview of the medgenvidqa 2026 shared task on medical generative video question answering. In Proceedings of the 25th Workshop on Biomedical Language Processing (BioNLP 2026), San Diego, USA. Association for Comput… find
  4. Deepak Gupta and Dina Demner-Fushman. 2022. Overview of the MedVidQA 2022 shared task on medical video question-answering. In Proceedings of the 21st Workshop on Biomedical Language Processing, pages 264–274, Dublin, Ireland. Association for Computational Linguistics. find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tian-2026-seahawk-medgenvidqa-2026-llm-segment/