**BiomedRAG: A Retrieval-Augmented Large Language Model for Biomedicine** **Authors:** Mingchen Li, Halil Kilicoglu, Hua Xu, and Rui Zhang **Abstract:** Large Language Models (LLMs) have become essential in biomedical and healthcare applications, but they often generate inaccurate information or hallucinations. Retrieval-augmented generation addresses these issues by updating knowledge and enhancing performance. Unlike previous retrieval-augmented LMs that use specialized cross-attention mechanisms, BiomedRAG directly inputs retrieved chunk-based documents into the LLM, bypassing noise information. This approach is easily applicable to existing models and demonstrates superior performance in four biomedical NLP tasks across eight datasets. Specifically, BiOMEDRAG retrieves relevant documents from a curated chunk database using a tunable scorer, integrating selected information into the LLM's input to improve predictions. Experiments show that BiOMEDRAG outperforms other triple extraction systems with micro-F1 scores of 81.42 and 88.83 on GIT and ChemProt corpora, respectively, highlighting its potential in constructing effective biomedical intervention systems. **Introduction:** The volume of biomedical literature has grown exponentially, necessitating advanced data mining and statistical techniques. Biomedical Large Language Models (LLMs) provide support for medical professionals and enhance BioNLP systems through pre-training or fine-tuning on biomedical data. However, these models are prone to hallucination. Retrieval-augmented language models can reduce hallucination and increase knowledge coverage by retrieving external knowledge. Previous methods often require a fixed retriever, such as K-nearest neighbors (KNN), which performs knowledge retrieval on unlabelled sentences. In contrast, BiomedRAG integrates chunk knowledge into LLMs, using a tailored chunk scorer to adapt the LLM and prioritize retrieving chunk-based documents to enhance language model perplexity. **Results:** BiomedRAG demonstrates superior performance in triple extraction, relation extraction, text classification, and link prediction tasks. For example, on the triple extraction task, BiOMEDRAG outperforms other systems with micro-F1 scores of 81.42 and 88.83 on GIT and ChemProt corpora, respectively. The model's effectiveness is further validated through ablation studies and comparisons with other models and retrieval-based LLMs. **Discussion:** BiomedRAG enhances the performance of LLMs by integrating retrieved chunk documents, leveraging a tailored chunk scorer to adapt the LLM and improve diversity. The model's robustness and versatility are demonstrated through experiments on various datasets, showing that it can handle noise-intensive tasks and datasets effectively. **Conclusion:** BiomedRAG is a novel biomedical RAG framework that integrates knowledge from a diverse chunk database and adapts the tailored chunk scorer to the LLM. Experimental results show consistent improvements in four biomedical NLP tasks across eight datasets.**BiomedRAG: A Retrieval-Augmented Large Language Model for Biomedicine** **Authors:** Mingchen Li, Halil Kilicoglu, Hua Xu, and Rui Zhang **Abstract:** Large Language Models (LLMs) have become essential in biomedical and healthcare applications, but they often generate inaccurate information or hallucinations. Retrieval-augmented generation addresses these issues by updating knowledge and enhancing performance. Unlike previous retrieval-augmented LMs that use specialized cross-attention mechanisms, BiomedRAG directly inputs retrieved chunk-based documents into the LLM, bypassing noise information. This approach is easily applicable to existing models and demonstrates superior performance in four biomedical NLP tasks across eight datasets. Specifically, BiOMEDRAG retrieves relevant documents from a curated chunk database using a tunable scorer, integrating selected information into the LLM's input to improve predictions. Experiments show that BiOMEDRAG outperforms other triple extraction systems with micro-F1 scores of 81.42 and 88.83 on GIT and ChemProt corpora, respectively, highlighting its potential in constructing effective biomedical intervention systems. **Introduction:** The volume of biomedical literature has grown exponentially, necessitating advanced data mining and statistical techniques. Biomedical Large Language Models (LLMs) provide support for medical professionals and enhance BioNLP systems through pre-training or fine-tuning on biomedical data. However, these models are prone to hallucination. Retrieval-augmented language models can reduce hallucination and increase knowledge coverage by retrieving external knowledge. Previous methods often require a fixed retriever, such as K-nearest neighbors (KNN), which performs knowledge retrieval on unlabelled sentences. In contrast, BiomedRAG integrates chunk knowledge into LLMs, using a tailored chunk scorer to adapt the LLM and prioritize retrieving chunk-based documents to enhance language model perplexity. **Results:** BiomedRAG demonstrates superior performance in triple extraction, relation extraction, text classification, and link prediction tasks. For example, on the triple extraction task, BiOMEDRAG outperforms other systems with micro-F1 scores of 81.42 and 88.83 on GIT and ChemProt corpora, respectively. The model's effectiveness is further validated through ablation studies and comparisons with other models and retrieval-based LLMs. **Discussion:** BiomedRAG enhances the performance of LLMs by integrating retrieved chunk documents, leveraging a tailored chunk scorer to adapt the LLM and improve diversity. The model's robustness and versatility are demonstrated through experiments on various datasets, showing that it can handle noise-intensive tasks and datasets effectively. **Conclusion:** BiomedRAG is a novel biomedical RAG framework that integrates knowledge from a diverse chunk database and adapts the tailored chunk scorer to the LLM. Experimental results show consistent improvements in four biomedical NLP tasks across eight datasets.