Reproduce “LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion” on OpenResearch

About this paper

Dongfu Jiang, Xiang Ren, Bill Yuchen Lin

The LLM-BLENDER framework, developed by researchers from Zhejiang University, USC, and AI2, introduces a two-module approach combining pairwise ranking and generative fusion to ensemble diverse open-source Large Language Models. This method consistently achieves superior performance, reducing the average GPT-Rank to 3.01 and ranking its outputs in the top 3 for 68.59% of examples on the newly released MixInstruct benchmark.

Key insights

• No single open-source LLM consistently achieves optimal performance across diverse inputs, validating the need for dynamic ensembling strategies.
• Pairwise comparison models (PAIRRANKER) are more effective at discerning subtle quality differences between high-quality LLM outputs than traditional pointwise ranking methods.
• Generative fusion of top-ranked candidates can produce outputs superior to merely selecting the best individual response, leveraging complementary strengths for synergistic improvement.

Summary by alphaXiv.