Authors: Anderson Morillo, Carlos Agamez, Edwin Puertas, Juan Carlos Martinez-Santos, Jairo Serrano
Venue: 2025 IEEE Colombian Caribbean Conference (C3)
DOI: 10.1109/C366505.2025.11340582
Links: IEEE Xplore
Full PDF text was not available for automated extraction (IEEE access). The summary below is grounded in the official published abstract and metadata.
Overview
This conference paper presents a resource-efficient biomedical question-answering architecture. Instead of relying on heavyweight dense retrieval stacks alone, it combines LLM-based key-term extraction and statistical weirdness scoring with document retrieval through the PubMed API, then generates answers with prompt engineering and RAG using Qwen3-30B-A3B.
Problem
Biomedical QA systems often trade accuracy for large compute budgets in retrieval and generation. The authors target a lighter pipeline that still uses structured literature access (PubMed) and modern LLMs, aiming at deployments where computational cost matters.
Method
According to the official abstract, the system:
- Extracts key terms with large language models
- Scores terms with statistical weirdness relative to a general corpus
- Retrieves literature via the PubMed API
- Generates answers with few-shot and instructional prompting plus Retrieval-Augmented Generation (RAG) using the Qwen3-30B-A3B model
Evaluation is reported on the BioASQ benchmark.
Results
The abstract states competitive performance on Factoid and Summary question types on BioASQ. Exact numeric scores are not restated here because the full PDF was not accessible for verification; see the IEEE record for tables and experimental detail.
Why it matters
- Connects the lab’s earlier weirdness line of work to biomedical retrieval
- Emphasizes a lightweight hybrid design (API retrieval + LLM RAG) rather than only high-compute dense search
- Positions the stack for practical biomedical QA where PubMed remains the literature source of truth