2026-04-14T02:27:14Zhttps://recercat.cat/oai/request

oai:recercat.cat:2117/4139692025-07-22T23:56:29Zcom_2072_1033col_2072_452951

00925njm 22002777a 4500 dc Bayarri Planas, Jordi author 2024-06-27 Foundational models have rapidly emerged in recent years, demonstrating remarkable capabilities across a wide array of tasks, predominantly in natural language processing. Significant efforts have been dedicated to this field, resulting in the frequent release of new, increasingly sophisticated models. This thesis explores the efficacy of advanced prompting strategies applied to these foundational models within the medical question-answering domain, focusing on the potential of open-source models enhanced through sophisticated prompt engineering. An efficient and functional evaluation framework, named "prompt\_engine", has been developed to study the potential of two prompting strategies: Self-Consistency Chain of Thought and a Medprompt-based technique. Through this framework, a comprehensive range of experiments was conducted, leading to significant performance enhancements through the strategic combination and optimization of these prompting techniques. Key findings reveal that the performance of open-source models can be significantly enhanced, allowing them to outperform current state-of-the-art private models in existing medical question-answering benchmarks. Àrees temàtiques de la UPC::Informàtica::Llenguatges de programació Natural language processing (Computer science) Models fundacionals models de llenguatge engineyria d'instruccions avaluació mèdica Foundational models large language models prompt engineering medical benchmarks Tractament del llenguatge natural (Informàtica) Prompt engineering for medical foundational models