When to Ensemble: Identifying Token-Level Points for Stable and Fast LLM Ensembling

Researchers have developed SAFE, a new framework for ensembling Large Language Models that selectively combines models at specific token positions rather than every token. The method improves both accuracy and efficiency in long-form text generation by considering tokenization mismatches and consensus in probability distributions.

• →Traditional LLM ensembling methods that combine models at every token often degrade performance in long-form generation tasks.
• →SAFE framework identifies optimal ensembling positions by analyzing tokenization mismatches and consensus in next-token probability distributions.
• →The method includes probability sharpening to improve stability when ensemble distributions become too smooth.
• →SAFE achieves performance gains while ensembling fewer than 1% of tokens, significantly improving efficiency.
• →Experiments on MATH500 and BBH benchmarks demonstrate superior accuracy and efficiency compared to existing ensemble methods.