paper
KR Labs at ArchEHR-QA 2025: A Verbatim Approach for Evidence-Based Question Answering
Abstract
Retrieval-augmented generation systems remain prone to fabrication. This paper presents a lightweight, domain-agnostic verbatim pipeline for evidence-grounded question answering: an extractor selects relevant source sentences, and an LLM drafts a question-specific template that is filled verbatim with extracted evidence. In the ArchEHR-QA 2025 shared task, the system scored 42.01%, ranked top-10 in core metrics, and outperformed the organiser's 70B Llama-3.3 baseline.
TL;DR
- The verbatim pipeline constrains answer content through extraction + templating.
- The paper reports a 42.01% ArchEHR-QA score, top-10 in core metrics.
- The system outperformed the organiser's 70B Llama-3.3 baseline.
- Open-source library at github.com/KRLabsOrg/verbatim-rag.
Resources
For the full text, citation list, and supplementary material, follow the paper link above. The library, models, and examples live in the open-source repository.
Cite
@inproceedings{kovacs-etal-2025-kr,
title = {{KR} Labs at {A}rch{EHR}-{QA} 2025: A Verbatim Approach for Evidence-Based Question Answering},
author = {Kovacs, Adam and Schmitt, Paul and Recski, Gabor},
booktitle = {Proceedings of the 24th Workshop on Biomedical Language Processing (Shared Tasks)},
year = {2025},
address = {Vienna, Austria},
publisher = {Association for Computational Linguistics},
pages = {69--74},
url = {https://aclanthology.org/2025.bionlp-share.8/},
doi = {10.18653/v1/2025.bionlp-share.8}
}