LLM-Blender is a research framework for ensembling multiple open-source LLMs to achieve consistently superior performance. It uses PairRM, a pairwise reward model, to rank outputs from different LLMs and identifies their respective strengths and weaknesses. A generative fusion module then combines the best aspects of each model to enhance capability. Published at ACL 2023 by researchers from AI2-Mosaic and USC INK lab, with models and datasets available on Hugging Face.