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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research, Natural Language Processing · Depth: Expert, medium

Summary

Xiaotian Tian and Gulustan Dogan's paper, "Seahawk at MedGenVidQA 2026," introduces a novel retrieval-and-selection pipeline for medical visual answer localization, a task requiring identification of temporal spans in videos that answer medical questions. Their system, designed for Task C, frames this as selecting contiguous transcript segments from timestamped video transcripts. Instead of directly generating timestamps, the system prompts DeepSeek to choose a segment range, from which start and end times are deterministically calculated. This method significantly decreases the risk of hallucinated or malformed temporal outputs. To handle extended video content, the approach employs overlapping sliding-window prompting and ranks potential ranges using lexical question matching. A 20-sample sanity check on a test dataset demonstrated that a completeness-biased configuration achieved an mIoU of 0.3217, while a duration-penalized configuration improved performance to 0.4815. These results position the constrained LLM-based segment selection with deterministic timestamp extraction as a practical baseline.

Key takeaway

For Machine Learning Engineers developing medical visual answer localization systems, you should consider adopting a constrained LLM segment selection approach. This method, which deterministically extracts timestamps from selected transcript segments, significantly mitigates hallucination risks common with direct timestamp generation. Implement overlapping sliding-window prompting to efficiently process long video transcripts, improving robustness and accuracy in your applications.

Key insights

Constrained LLM segment selection with deterministic timestamp extraction effectively localizes medical video answers, reducing hallucination risks.

Principles

• Deterministic timestamp extraction prevents LLM hallucination.
• Segment-level selection simplifies temporal localization.
• Overlapping windows handle long video transcripts.

Method

A retrieval-and-selection pipeline prompts DeepSeek to select contiguous transcript segments. Start/end times are deterministically computed from segment boundaries. Overlapping sliding-window prompting and lexical ranking support long videos.

In practice

• Use DeepSeek for segment range selection.
• Implement deterministic timestamp calculation.
• Apply sliding-window prompting for long inputs.

Topics

• Medical Visual QA
• LLM Segment Selection
• DeepSeek
• Timestamp Localization
• Video Transcripts
• BioNLP 2026

Best for: NLP Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• 5 Practical Techniques to Detect and Mitigate LLM Hallucinations Beyond Prompt Engineering — LLM hallucinations are a system problem requiring multi-layered detection and mitigation beyond prompt engineering.
• The Evolution of LLM Inference: Decoding algorithms — Part 1 — LLM inference speed hinges on reducing sequential decoding steps by predicting and verifying multiple tokens in parallel.
• When Cases Get Rare: A Retrieval Benchmark for Off-Guideline Clinical Question Answering — LLMs require retrieval augmentation and strong reasoning for reliable off-guideline medical question answering.
• FBK's Long-form SpeechLLMs for IWSLT 2026 Instruction Following — Fixed 30-second segmentation provides robust long-form SpeechLLM performance, mitigating hallucination effects on ASR and SSUM.
• Towards Generalization of Block Attention via Automatic Segmentation and Block Distillation — Semantic segmentation and an efficient distillation framework enable practical, scalable block attention for long-context LLMs.
• We Tested 25 Local LLMs for Medical Use. Here’s What Shipped. — Local LLMs are viable for medical AI, but require careful selection and a multi-model, multi-stage architecture.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.