Reproduce “Cool-Fusion: Fuse Large Language Models without Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cong Liu, Xiaojun Quan, Yan Pan, Liang Lin, Weigang Wu, Xu Chen
A training-free approach for fusing heterogeneous Large Language Models (LLMs) is introduced, operating at the text segment level to bypass architectural and vocabulary discrepancies. This method, called Cool-Fusion, demonstrated up to a 17.4% accuracy improvement over the best individual LLM on the GSM8K benchmark.
Key insights
- Fusing Large Language Models at the text segment level bypasses token-level and architectural incompatibilities, enabling heterogeneous model combinations.
- Averaged perplexity across all source LLMs provides a training-free and robust metric for jointly evaluating the quality and coherence of generated text segments.
- Combining fine-grained iterative segment selection with coarse-grained full-continuation reranking leverages both local context improvement and global output evaluation for enhanced performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.