When to Ensemble: Identifying Token-Level Points for Stable and Fast LLM Ensembling
This addresses the challenge of stable and efficient ensembling for long-form LLM generation, which is an incremental improvement over existing methods.
The paper tackled the problem of applying LLM ensembling to long-form generation by identifying that ensembling at every token degrades performance, and proposed SAFE, a framework that selectively ensembles based on tokenization mismatch and probability consensus, achieving gains in accuracy and efficiency with ensembling on fewer than 1% of tokens.
Ensembling Large Language Models (LLMs) has gained attention as a promising approach to surpass the performance of individual models by leveraging their complementary strengths. In particular, aggregating models' next-token probability distributions to select the next token has been shown to be effective in various tasks. However, while successful for short-form answers, its application to long-form generation remains underexplored. In this paper, we show that using existing ensemble methods in long-form generation requires a careful choice of ensembling positions, since the standard practice of ensembling at every token often degrades performance. We identify two key factors for determining these positions: tokenization mismatch across models and consensus in their next-token probability distributions. Based on this, we propose SAFE, (Stable And Fast LLM Ensembling), a framework that selectively ensembles by jointly considering these factors. To further improve stability, we introduce a probability sharpening strategy that consolidates probabilities spread across multiple sub-word tokens representing the same word into a single representative token. Our experiments on diverse benchmarks, including MATH500 and BBH, demonstrate that SAFE outperforms existing methods in both accuracy and efficiency, with gains achieved even when ensembling fewer than 1% of tokens.