This paper presents an overview of the Question Answering (QA) tasks in the thirteenth edition of the BioASQ challenge, on large-scale biomedical semantic indexing and QA, which is part of the Conference and Labs of the Evaluation Forum (CLEF) 2025. For more than a decade, BioASQ has been serving as a key platform for advancing the state-of-the-art in biomedical information retrieval, NLP, and QA. In this paper, we present a comprehensive overview of the biomedical QA tasks 13b and Synergy13 of the thirteenth BioASQ challenge (BioASQ 13). This year, 49 teams with more than 160 systems participated in these two tasks of the challenge, with 46 focusing on task 13b, on biomedical semantic QA, and 5 on task Synergy13, on QA for open questions on developing biomedical topics. The competitive performance achieved by several participating systems in the QA tasks of BioASQ 13 highlights the continuous advancement of state-of-the-art methods in the field, in alignment with previous editions of the tasks.
Task 13b introduces a comprehensive question-answering challenge in the biomedical field, requiring
participants to develop systems that address all stages of question answering. As in previous editions,
the task focuses on four question types: ‘yes/no,’ ‘factoid,’ ‘list,’ and ‘summary’ questions [
In the thirteenth edition of the BioASQ Challenge, participating teams received an updated version of
the BioASQ QA training dataset [
Task 13b, similar to the previous version of the task (12b), was structured into three phases [
Introduced in the ninth BioASQ edition [13], the Synergy task aimed to foster collaboration between biomedical experts studying COVID-19 and automated question-answering systems participating in BioASQ. The core objective is to create a synergy where experts assess system responses, and this feedback is used to iteratively improve the systems [14]. The task continued to focus on COVID-19 in the tenth edition of BioASQ [15], and it was extended to any developing biomedical topic in the subsequent editions of the challenge [16, 17, 18].
In the process depicted in Figure 1, competing systems provide their initial responses to open questions related to emerging problems. These responses, along with relevant documents and snippets, are evaluated by experts. Subsequently, the experts provide feedback to the systems and address any new or pending questions.
As in previous years, the Synergy task (Synergy13) comprised four rounds, with a two-week interval
between each round, and focused on emerging biomedical topics, drawing from relevant documents
in the current PubMed version [
In the Synergy task, during each round, the system responses and expert feedback address the same
questions, unless those questions have already been closed by experts due to receiving a comprehensive
and definite answer. Specifically, in Synergy13, a group of four biomedical experts contributed a total of
74 open biomedical questions. This set includes 47 new questions on developing health topics, such as
infectious, rare, and genetic diseases, and women’s and reproductive health, and 27 questions from the
previous version of the task [
Synergy13, similar to task 13b, addresses four question types: yes/no, factoid, list, and summary, and two types of answers, exact and ideal. Moreover, the evaluation of systems relies on the same measures used in task 13b. Upon completing the Synergy13 task, relevant material was identified for answering roughly 80% of the questions. Additionally, around 47% of the questions had at least one ideal answer submitted by the systems, which was deemed satisfactory by the expert who posed the question.
In this year’s BioASQ challenge, over 160 distinct systems engaged in tasks 13b and Synergy13 with a total of 49 teams, 39 of which were new. The high percentage of teams joining the BioASQ challenge for the first time, which is almost 80%, indicates the enduring interest of the community in large-scale biomedical question answering. Specifically, 46 of these teams submitted on at least some phase of task 13b and 5 on task Synergy13, with several teams participating in more than one task and phases, as demonstrated in Figure 2.
In line with previous years [
An overview of the technologies utilized by the teams is outlined in Table 3. Additional details for specific systems can be found in the workshop proceedings. As in previous years, the open-source system OAQA [21], which achieved top performance in older editions of BioASQ [22, 23], was used as a baseline for phase B exact answers. This system is based on the UIMA framework and relies on UA UR traditional NLP and Machine Learning approaches and tools, such as MetaMap and LingPipe 1.
The UA team from the University of Aveiro participated in all three phases of the task with five systems. In phase A, their systems followed a two-stage retrieval pipeline in line with previous years’ submissions [41, 42]. To enhance the BM25 results they used the BGE-M3 model, and reciprocal rank fusion to combine model outputs. For phases A+ and B, their systems adapted RAG with prompts incorporating relevant abstracts using instruction-based transformer models such as llama-3 70B and Gemma-3 27B.
The UR team from the Universität Regensburg competed in all phases of the task with five systems. Their systems employed few-shot prompting (typically 0-shot or 10-shot configurations) and an experimental two-step self-feedback mechanism. In this feedback approach, an LLM generated an initial answer, then provided critical feedback on its own output, and finally refined the answer based on this critique. This process was applied across yes/no, factoid, list, and ideal answer types. For the retrieval stages, their systems utilized LLM-driven query expansion, which also incorporated a similar self-feedback loop for query refinement based on initial search results, coupled with Elasticsearch for document retrieval, followed by LLM-based snippet extraction and re-ranking.
The NCU team from the National Central University participated with five systems. Their systems employed a basic Retrieval-Augmented Generation (RAG) framework consisting of a retriever, a reranker, and a LLM for language generation. The initial retriever used was a BM25, and the documents were further re-ranked using the bge-eranker-v2-m3 model to identify the most relevant articles and snippets. For answer generation, their systems used the meta-llama/Llama-3.1-8B-Instruct and GPT-4o modles. Furthermore, for the Phase A+ task, they extended the answer generation pipeline previously developed by Chih et al. [43]. For phase B, the systems utilized GPT-4o, o3-mini, and Gemini 2.0 Flash. Each system employed either direct or two-stage generation for ideal and exact answers.
Another team participating in all phases is the team from the BSRC Alexander Fleming Institute. For Phases A and A+, their systems focused on document retrieval using the BM25 algorithm enhanced with RM3, and re-ranking based on the relevance of associated text snippets, similar to the previous year participation [44]. In phases A+ and B, their systems explored three distinct prompting strategies for generating exact answers (snippet-based, abstract-based, and extended abstract-based). To further enhance answer quality, their systems build upon their earlier LLM ensemble strategy [44] and extend it to all exact answer types. For list and factoid questions, they use the union of answers across multiple LLMs to maximize recall.
The PJs team participated in all three phases, submitting five system configurations in total. For Phase A, their systems employed dense retrieval using the e5-base model and then re-ranking using diferent models (e.g. modernbert-embed-base, gte-large, granite-25m, gte ensemble, modernbert, and granite). For snippet selection, MiniLM-L6 was utilized. For Phase A+, their systems used Claude Sonnet 3 with prompt engineering on the top retrieved results, while for Phase B, the the same prompt design was utilized with diferent LLM modles including Qwen2.5-3B, Claude Sonnet 3.7, Claude Haiku 3.5, and an ensemble of Sonnet 3.7 and Sonnet 3.0 with post-processing to correct errors.
Also, the JU_NLP team from Jadavpur University participated in all phases. For phase A, their systems perform dense retrieval using a fine-tuned minilm model. Then, a RoBERTa-base-SQuAD2 model extracts the answer snippet. For phase B, a separately fine-tuned T5 model for generating ideal and exact answers is utilized. Finally, phase A+ combines these processes, first using the Phase A system to extract relevant articles and snippets, and then feeding those results into the Phase B system to produce ideal and exact answer.
The UniTor team from Università degli Studi di Roma Tor Vergata participated in all three phases. Their systems employed a multi-stage pipeline that combined sparse and dense retrieval techniques, supervised re-ranking, supervised LLM snippet extraction, and supervised LLM answer generation. All components were trained and evaluated using oficial BioASQ datasets and external biomedical resources such as PubMed.
The UniBol team from Universidad Tecnológica de Bolivar participated in all three phases. Their systems employed the PubMedBert model to perform dense retrieval and Qwen3 8b with prompt engineer to generate answers.
Another team that participated in all phases is the lasigeBioTM team. Their systems used Mistral as the baseline model for all phases. Furthermore, they incorporated external knowledge from various ontologies and knowledge bases like Human Disease Ontology and NCBI Gene, with BENT tool used for named-entity recognition and linking. Also, some systems employed a Decoding on Graphs methodology, utilizing Monarch KG and MARISA Trie for structured reasoning.
The AQAMS team from Universidad Europea took part in all phases. For phases A and A+ their systems used a two-stage pipeline. First, LlamaIndex embeddings were used to perform dense retrieval. Then, OpenAI GPT models with structured prompt templates were utilized to generate answers. For phase B, their systems focused on biomedical NER using a fine-tuned model, optionally mapping terms to UMLS concepts, and generating answers, all deployed via a user-friendly Streamlit interface.
The ZUT team from Zhongyuan University of Technology participated in all phases. For phase A, their systems employed a query expansion-driven multi-stage retrieval and re-ranking framework that utilized diferent DeepSeek models. In phases A+ and B, their systems utilized diferent prompting strategies and also experimented with supervised fine-tuning of LLMs.
The FSU team from Florida State University participated in phases A and A+. Their systems followed a multi-stage retrieval and reasoning framework, leveraging an LLM for keyword extraction and answer generation, along with an internal search engine for document retrieval. A re-ranker model and an LLM-based scoring mechanism filter retrieved documents for quality. Additionally, their systems integrate structured biomedical knowledge from IKraph, an in-house knowledge graph built from PubMed abstracts using large-scale relation extraction and causal inference.
The GT team from Georgia Tech participated in both phases A and B. For phase A, their systems employed a custom index built with the bge-large-en embedding model and used a fine-tuned ms-marco12 model for re-ranking, with some configurations also incorporating GPT-4o as an additional re-ranker. In phase B, the team utilized Mistral-7B-Instruct-v0.3 and GPT-4o-turbo, with a prompt-based approach using few-shot examples to generate answers.
The MQU team from Macquarie University participated in phases A+ and B. Their systems employed a zero-shot QA framework, using prompting multiple LLMs (Gemini variants and Claude) to generate answers based on snippets and full abstracts. A secondary synthesis step with confidence-based re-tries refines the candidate answers into a final response, improving precision and consistency across yes/no and factoid questions.
Also, the HCMC team from the University of Science participated in phases A+ and B. Their systems ifrst distinguish queries as single-hop or multi-hop; multi-hop queries are decomposed into sub-queries. Relevant documents are retrieved using the PubMed API, with queries reformulated into Conjunctive Normal Form (CNF). Document sentences are encoded using SciSpaCy, and dense retrieval is performed using all-MiniLM-L12-v2 model. Finally, GPT-4o-mini is utilized to generate answers.
The VCU team from Virginia Commonwealth University participated with four diferent systems in phase B. Their systems are based on a zero-shot learning approach using generative LLMs, including Synthia-13B-GPTQ and llama-3. Their systems heavily relied on prompt engineering and answer processing.
The Evidence team from Evidence Prime participated in phase B. Their systems employed dense retrieval using the nomic-embed-text-v1 model. Various open-source LLMs and proprietary models were used, including GPT-4o, GPT-4.1, Claude 3.5, and Claude 3.7. For Yes/No and Factoid/List questions, an ensemble strategy was implemented, using voting or frequency analysis of outputs from multiple LLMs. Summary responses were generated using carefully designed, hand-crafted prompts to balance contextual similarity and appropriate length.
The DMIS team from the Korea University participated in phase B. Their systems employed multiple LLMs, including GPT-4o-mini, GPT-4, and Claude. They explored various prompting strategies, such as standard instruction, one-by-one snippet querying, randomized snippet order, and a no-snippet condition relying on prior knowledge. Final answers were derived through ensemble techniques, either by aggregating log probability scores or using majority voting.
Another team that participated only in phase B is the UTB team from the Universidad Tecnológica de Bolívar. Their systems were based on the well-known BERT model.
In the thirteenth edition of BioASQ, five teams participated in the Synergy task (Synergy13). These teams submitted 46 runs from 21 distinct systems. Two of these teams participated in task 13b as well, while the remaining three focused exclusively on task Synergy13. An overview of systems and approaches employed is provided in Table 4.
In particular, the BSRC Alexander Fleming team participated with four systems. Similar to task b, their systems focused on LLMs, specifically the DeepSeek-R1 model, with optimized prompts and majority voting. Also, the SINAI team from the Universidad de Jaén competed with five systems. Their systems built upon prior research by integrating a lightweight NER-based query module, dynamic indexing of PubMed abstracts, and few-shot prompt templates tailored to diferent question types. These components are utilized within a RAG framework based on a biomedical fine-tuned LLaMA model. The system employs a multi-stage pipeline—comprising data preparation, context extraction, and response generation—designed specifically for biomedical question answering across the required formats (summary, yes/no, factoid, list).
4.1. Task 13b This section presents the evaluation measures and preliminary results for the task 13b. These results are preliminary, as the final results will be available after the manual assessment of the system responses by the BioASQ team of experts and the enrichment of the ground truth with potential additional relevant items, answer elements, and/or synonyms, which is still in progress.
Phase A: The Mean Average Precision (MAP) was used for evaluation on document retrieval. In particular, since BioASQ8 [46], MAP calculation is based on a modified version of Average Precision (AP) that considers both the limit of 10 elements allowed per question in each submission and the actual number of golden elements that is often less than 10 in practice [47]. For snippets, where a single ground-truth snippet may overlap with several submitted ones, the interpretation of MAP is less straightforward. Hence, since BioASQ9 [13], we use the F-measure which is based on character overlaps2 [48]. Tables 5 and 6 present some indicative results in batch 1 for document and snippet retrieval, respectively. The full 13b results for phase A are available online3.
Phases A+ and B: The oficial ranking for systems providing ideal answers is based on manual scores
assigned by the BioASQ team of experts that assesses each ideal answer in the responses [
The top performance of the participating systems in exact answer generation for each type of question during the thirteen years of BioASQ is presented in Figure 4. The preliminary results for task 13b, reveal that the participating systems keep achieving high scores in answering all types of questions, despite the addition of two new experts to the BioASQ team. In batch 2, phase B, for instance, presented in Table 8, several systems manage to correctly answer literally all yes/no questions, as well as in batches 2http://participants-area.bioasq.org/Tasks/b/eval_meas_2022/ 3http://participants-area.bioasq.org/results/13b/phaseA/ 4http://participants-area.bioasq.org/results/13b/phaseAplus/ 5http://participants-area.bioasq.org/results/13b/phaseB/ 1 and 4. In phase A+, on the other hand, where no ground truth relevant material is available, correctly answering all yes/no questions is more challenging, though not infeasible. In batch 2, for instance, presented in Table 7, only one system manages to do so. The preliminary results of task 13b, Phase B, reveal improved and more consistent performance for list questions compared to the previous years, but room for improvement is still available, as is for factoid questions, where the performance across batches presents more fluctuations.
In task Synergy13, no relevant material was initially available for new questions. For old questions, however, feedback from previous rounds was provided per question, that is the documents and snippets submitted by the participants with manual annotations of their relevance. Hence, the documents and snippets of the feedback, that have already been assessed and released, were not considered valid for submission in the subsequent rounds. As in task 13b, the evaluation measures for document and snippet retrieval are MAP and F-measure respectively.
In addition, due to the developing nature of the topic, no answer is available for all of the open
questions in each round. Therefore only the questions indicated as “answer ready” were evaluated for
exact and ideal answers in each round. Regarding the ideal answers, the systems were ranked according
to manual scores assigned to them by the BioASQ experts during the assessment of systems responses
as in phase B of task B [
Some indicative results for the Synergy task are presented in Table 9. The full Synergy13 results are available online6. Overall, the collaboration between participating biomedical experts and questionanswering systems allowed the progressive identification of relevant material and extraction of exact and ideal answers for several open questions for developing problems, such as infectious, rare, and genetic diseases, and women’s and reproductive health. In total, after the four rounds of Synergy13, enough relevant material was identified to provide an answer to about 80% of the questions. In addition, about 47% of the questions had at least one ideal answer, submitted by the systems, which was considered of ground-truth quality by the respective expert. 6http://participants-area.bioasq.org/results/synergy_v2025/
In this paper, we introduced the thirteenth version of the BioASQ challenge, focusing on the question answering tasks b and Synergy. These tasks have been well-established through previous versions of the challenge and remain timely and relevant, as indicated by the increased participation.
The preliminary results of task 13b reveal the strong performance of top participating systems, particularly in generating yes/no answers, even under the constraints of Phase A+, where no groundtruth relevant documents and snippets were provided. System performance on list and factoid questions was more variable, especially in Phase A+, highlighting the presence of room for improvement. These results suggest that access to ground truth relevant material significantly enhances response quality for more complex question types. This emphasizes the critical role of Phase A, which involves the automatic retrieval of relevant documents and snippets for answering a biomedical question. Performance in Phase A showed greater variability across batches, potentially influenced by the domain expertise of the experts who authored the questions. A diverse set of retrieval and generation strategies was employed, including traditional IR methods, large language model (LLM)-based approaches, and systems enriched with domain-specific biomedical knowledge. Finally, the results of Synergy13 highlight that state-of-the-art QA systems can be useful in aiding biomedical researchers to address their specialized information needs, in alignment with the results of previous versions of the task, despite the persisting challenges and room for improvement.
Overall, several participating systems achieved competitive performance on the QA tasks of BioASQ 13, and some of them managed to improve over the state-of-the-art performance from previous years. Therefore, twelve years after its initial introduction, BioASQ keeps pushing the research frontier in biomedical question answering, ofering two QA tasks and several response types that cover a range of biomedical information needs.
The thirteenth edition of BioASQ is sponsored by Ovid, Atypon Systems Inc, and Elsevier. The MEDLINE/PubMed data resources considered in this work were accessed courtesy of the U.S. National Library of Medicine. BioASQ is grateful to the CMU team for providing the exact answer baselines for task 13b.
During the preparation of this work, the authors used Grammarly and ChatGPT to perform the following tasks: grammar and spelling checks, Paraphrasing, and rewording. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [13] BioASQ at CLEF2021: Large-Scale Biomedical Semantic Indexing and Question Answering, in: Advances in Information Retrieval, Springer International Publishing, Springer International Publishing, Cham, 2021. [14] A. Krithara, A. Nentidis, E. Vandorou, G. Katsimpras, Y. Almirantis, M. Arnal, A. Bunevicius, E. Farré-Maduell, M. Kassiss, V. Konstantakos, S. Matis-Mitchell, D. Polychronopoulos, J. RodriguezPascual, E. G. Samaras, M. Samiotaki, D. Sanoudou, A. Vozi, G. Paliouras, BioASQ Synergy: a dialogue between question-answering systems and biomedical experts for promoting COVID19 research, Journal of the American Medical Informatics Association (2024) ocae232. URL: https://doi.org/10.1093/jamia/ocae232. doi:10.1093/jamia/ocae232. [15] A. Nentidis, A. Krithara, G. Paliouras, L. Gasco, M. Krallinger, BioASQ at CLEF2022: The Tenth Edition of the Large-scale Biomedical Semantic Indexing and Question Answering Challenge, in: Advances in Information Retrieval: 44th European Conference on IR Research, ECIR 2022, Stavanger, Norway, April 10–14, 2022, Proceedings, Part II, Springer, Springer, 2022. URL: https: //link.springer.com/chapter/10.1007/978-3-030-99739-7_53. [16] A. Nentidis, A. Krithara, G. Paliouras, E. Farre-Maduell, S. Lima-Lopez, M. Krallinger, BioASQ at CLEF2023: The Eleventh Edition of the Large-Scale Biomedical Semantic Indexing and Question Answering Challenge, in: Advances in Information Retrieval: 45th European Conference on Information Retrieval, ECIR 2023, Dublin, Ireland, April 2–6, 2023, Proceedings, Part III, Springer, 2023, pp. 577–584. [17] A. Nentidis, A. Krithara, G. Paliouras, M. Krallinger, L. G. Sanchez, S. Lima, E. Farre, N. Loukachevitch, V. Davydova, E. Tutubalina, BioASQ at CLEF2024: The Twelfth Edition of the Large-Scale Biomedical Semantic Indexing and Question Answering Challenge, in: European Conference on Information Retrieval, Springer, 2024, pp. 490–497. [18] A. Nentidis, G. Katsimpras, A. Krithara, M. Krallinger, M. R. Ortega, N. Loukachevitch, A. Sakhovskiy, E. Tutubalina, G. Tsoumakas, G. Giannakoulas, A. Bekiaridou, A. Samaras, G. M. Di Nunzio, N. Ferro, S. Marchesin, L. Menotti, G. Silvello, G. Paliouras, BioASQ at CLEF2025: The Thirteenth Edition of the Large-Scale Biomedical Semantic Indexing and Question Answering Challenge, in: Advances in Information Retrieval, Springer Nature Switzerland, Cham, 2025, pp. 407–415. [19] A. Nentidis, G. Katsimpras, E. Vandorou, A. Krithara, G. Paliouras, Overview of bioasq tasks 9a, 9b and synergy in clef2021, in: Proceedings of the 9th BioASQ Workshop A challenge on large-scale biomedical semantic indexing and question answering, 2021. URL: http://ceur-ws.org/Vol-2936/ paper-10.pdf. [20] A. Nentidis, G. Katsimpras, E. Vandorou, A. Krithara, G. Paliouras, Overview of BioASQ Tasks 10a, 10b and Synergy10 in CLEF2022, in: Proceedings of the 10th BioASQ Workshop A challenge on large-scale biomedical semantic indexing and question answering, 2022. URL: https://ceur-ws. org/Vol-3180/paper-10.pdf. [21] Z. Y. Y. Z. E. Nyberg, Learning to Answer Biomedical Questions: OAQA at BioASQ 4B, ACL 2016 (2016) 23. [22] G. Balikas, A. Kosmopoulos, A. Krithara, G. Paliouras, I. Kakadiaris, Results of the BioASQ tasks of the question answering lab at CLEF 2015, CEUR Workshop Proceedings 1391 (2015). [23] A. Krithara, A. Nentidis, G. Paliouras, I. Kakadiaris, Results of the 4th edition of BioASQ Challenge, in: Proceedings of the Fourth BioASQ workshop, Association for Computational Linguistics, Stroudsburg, PA, USA, 2016, pp. 1–7. URL: http://aclweb.org/anthology/W16-3101. doi:10.18653/ v1/W16-3101. [24] R. A. A. Jonker, T. Almeida, J. Almeida, S. Matos, BIT.UA at BioASQ 13B: Revisiting Evaluation, DPRF-Enhanced Retrieval and Fine-Tuned LLMs, in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [25] S. Ateia, U. Kruschwitz, Can Language Models Critique Themselves? Investigating Self-Feedback for Retrieval Augmented Generation at BioASQ 2025 , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [26] C. Bing-Chen, J.-C. Han, H.-C. Hung, R. T.-H. Tsai, NCU-IISR: Biomedical Question Answering via Gemini and GPT APIs in the BioASQ 13b Phase B Challenge , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [27] J.-C. Han, B.-C. Chih, H.-C. Hung, R. T.-H. Tsai, A Retrieval-Augmented Generation Approach for BioASQ 13b Phase A and A+ , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [28] D. Panou, A. Dimopoulos, M. Koubarakis, M. Reczko, Harnessing Collective Intelligence of LLMs for Robust Biomedical QA: A Multi-Model Approach , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [29] P. Vachharajani, Exploring Retrieval-Reranking and LLM-Based Answer Generation for Biomedical
QA , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [30] H. P. Gupta, R. Banerjee, LLMs for Biomedical NER , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [31] F. Borazio, D. Croce, R. Basili, UniTor at BioASQ 2025: Modular Biomedical QA with Synthetic Snippets and Multiple Task Answer Generation , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [32] A. Morillo, E. Puertas, J. C. M. Santos, J. S. Castaneda, C. A. Palomino, VerbanexAI at BioASQ 13B: PubMed API and LLM-Driven Hybrid Retrieval for Biomedical Question Answering , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [33] P. R. C. Lopes, S. I. R. Conceição, M. Fernandes, F. M. Couto, lasigeBioTM: A lean biomedical QA system empowered by structured knowledge , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [34] J. Angulo, V. Yeste, AQAMS and AQAMS2: Multi Agent Systems for Biomedical Question Answering , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [35] J. Tang, H. Yang, K. Xiong, H. Li, P. Quaresma, H. Yu, W. Zhang, M. Song, Y. Jiang, Applying DeepSeek to BioASQ Task 13B: Using Supervised Fine-Tuning and Few-Shot Learning , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [36] S. Verma, F. Jiang, X. Xue, Beyond Retrieval: Ensembling Cross-Encoders and GPT Rerankers with LLMs for Biomedical QA , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [37] D. Galat, D. Molla-Aliod, LLM Ensemble for RAG: Role of context length in zero-shot Question Answering for BioASQ Challenge , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [38] E. Quiñones, E. A. P. D. Castillo, Evaluation of a System for Generating Exact and Ideal Responses , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [39] D. Stachura, J. Konieczna, A. Nowak, Are Smaller Open-Weight LLMs Closing the Gap to Proprietary Models for Biomedical Question Answering? , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [40] H. Kim, H. Lee, Y. Cho, J. Park, J. Park, S. Park, Y. T. Chok, S. Baek, D. Lee, J. Kang, Prompting Matters: Snippet-Aware Strategies for Biomedical QA with LLMs in BioASQ 13b , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [41] T. Almeida, R. A. A. Jonker, R. Poudel, J. M. Silva, S. Matos, Bit. ua at bioasq 11b: Two-stage ir with synthetic training and zero-shot answer generation., in: G. Faggioli, N. Ferro, P. Galuščáková, A. García Seco de Herrera (Eds.), Working Notes of CLEF 2024 - Conference and Labs of the Evaluation Forum, 2024. [42] T. Almeida, R. A. Jonker, J. Reis, J. R. Almeida, S. Matos, Bit. ua at bioasq 12: From retrieval to answer generation, in: G. Faggioli, N. Ferro, P. Galuščáková, A. García Seco de Herrera (Eds.), Working Notes of CLEF 2024 - Conference and Labs of the Evaluation Forum, 2024. [43] B.-C. Chih, J.-C. Han, R. Tzong-Han Tsai, NCU-IISR: Enhancing Biomedical Question Answering with GPT-4 and Retrieval Augmented Generation in BioASQ 12b Phase B, in: G. Faggioli, N. Ferro, P. Galuščáková, A. García Seco de Herrera (Eds.), CLEF Working Notes, 2024. [44] D. Panou, A. Dimopoulos, M. Reczko, Farming Open LLMs for Biomedical Question Answering, in: G. Faggioli, N. Ferro, P. Galuščáková, A. García Seco de Herrera (Eds.), CLEF Working Notes, 2024. [45] S. D. Romero, L. A. Ureña-López, E. Martínez-Cámara, SINAI at CLEF 2025: A Multi-Stage RAG Pipeline for Biomedical Semantic Question Answering , in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), CLEF 2025 Working Notes, 2025. [46] M. Krallinger, A. Krithara, A. Nentidis, G. Paliouras, M. Villegas, BioASQ at CLEF2020: large-scale biomedical semantic indexing and question answering, in: European Conference on Information Retrieval, Springer, 2020, pp. 550–556. [47] A. Nentidis, A. Krithara, K. Bougiatiotis, M. Krallinger, C. Rodriguez-Penagos, M. Villegas, G. Paliouras, Overview of BioASQ 2020: The eighth BioASQ challenge on Large-Scale Biomedical Semantic Indexing and Question Answering, in: Experimental IR Meets Multilinguality, Multimodality, and Interaction Proceedings of the Eleventh International Conference of the CLEF Association (CLEF 2020), Thessaloniki, Greece, September 22–25, 2020, Proceedings, volume 12260, Springer, 2020. [48] A. Nentidis, G. Katsimpras, E. Vandorou, A. Krithara, L. Gasco, M. Krallinger, G. Paliouras, Overview of BioASQ 2021: The Ninth BioASQ Challenge on Large-Scale Biomedical Semantic Indexing and Question Answering, in: International Conference of the Cross-Language Evaluation Forum for European Languages, Springer, 2021, pp. 239–263. [49] A. Nentidis, A. Krithara, K. Bougiatiotis, G. Paliouras, I. Kakadiaris, Results of the sixth edition of the BioASQ Challenge, in: Proceedings of the 6th BioASQ Workshop A challenge on largescale biomedical semantic indexing and question answering, 1, Association for Computational Linguistics, Brussels, Belgium, 2018, pp. 1–10. URL: http://aclweb.org/anthology/W18-5301. doi:10. 18653/v1/W18-5301. [50] A. Nentidis, G. Katsimpras, E. Vandorou, A. Krithara, A. Miranda-Escalada, L. Gasco, M. Krallinger, G. Paliouras, Overview of BioASQ 2022: The Tenth BioASQ Challenge on Large-Scale Biomedical Semantic Indexing and Question Answering, in: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), volume 13390 LNCS, 2022, pp. 337–361. doi:10.1007/978-3-031-13643-6_22. arXiv:2210.06852.