LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion
This work addresses the challenge of leveraging diverse LLMs for more reliable AI outputs, though it is incremental as it builds on existing ensembling techniques.
The authors tackled the problem of inconsistent performance across different large language models (LLMs) by developing LLM-Blender, an ensembling framework that uses pairwise ranking and generative fusion to combine outputs from multiple models, resulting in significantly superior performance across various metrics compared to individual LLMs and baseline methods.
We present LLM-Blender, an ensembling framework designed to attain consistently superior performance by leveraging the diverse strengths of multiple open-source large language models (LLMs). Our framework consists of two modules: PairRanker and GenFuser, addressing the observation that optimal LLMs for different examples can significantly vary. PairRanker employs a specialized pairwise comparison method to distinguish subtle differences between candidate outputs. It jointly encodes the input text and a pair of candidates, using cross-attention encoders to determine the superior one. Our results demonstrate that PairRanker exhibits the highest correlation with ChatGPT-based ranking. Then, GenFuser aims to merge the top-ranked candidates, generating an improved output by capitalizing on their strengths and mitigating their weaknesses. To facilitate large-scale evaluation, we introduce a benchmark dataset, MixInstruct, which is a mixture of multiple instruction datasets featuring oracle pairwise comparisons. Our LLM-Blender significantly outperform individual LLMs and baseline methods across various metrics, establishing a substantial performance gap.