Ensembling different large language models (LLMs) to unleash their
complementary potential and harness their individual strengths is highly
valuable. Nevertheless, vocabulary discrepancies among various LLMs have
constrained previous studies to either selecting or blending completely
generated outputs. This limitation hinders the dynamic correction and
enhancement of outputs during the generation process, resulting in a limited
capacity for effective ensemble. To address this issue, we propose a novel
method to Ensemble LLMs via Vocabulary Alignment (EVA). EVA bridges the lexical
gap among various LLMs, enabling meticulous ensemble at each generation step.
Specifically, we first learn mappings between the vocabularies of different
LLMs with the assistance of overlapping tokens. Subsequently, these mappings
are employed to project output distributions of LLMs into a unified space,
facilitating a fine-grained ensemble. Finally, we design a filtering strategy
to exclude models that generate unfaithful tokens. Experimental results on
commonsense reasoning, arithmetic reasoning, machine translation, and
data-to-text generation tasks demonstrate the superiority of our approach
compared with individual LLMs and previous ensemble methods conducted on
complete outputs. Further analyses confirm that our approach can leverage
knowledge from different language models and yield consistent improvement.

通过词汇对齐方法，将不同的语言模型通过集成的方式进行整合，并在生成过程中动态修正和增强输出，以提高集成模型的效果和性能。

跨越 LLM 集成中不同词汇之间的鸿沟

Bridging the Gap between Different Vocabularies for LLM Ensemble

The construction and maintenance process of the UMLS (Unified Medical
Language System) Metathesaurus is time-consuming, costly, and error-prone as it
relies on (1) the lexical and semantic processing for suggesting synonymous
terms, and (2) the expertise of UMLS editors for curating the suggestions. For
improving the UMLS Metathesaurus construction process, our research group has
defined a new task called UVA (UMLS Vocabulary Alignment) and generated a
dataset for evaluating the task. Our group has also developed different
baselines for this task using logical rules (RBA), and neural networks (LexLM
and ConLM).
In this paper, we present a set of reusable and reproducible resources
including (1) a dataset generator, (2) three datasets generated by using the
generator, and (3) three baseline approaches. We describe the UVA dataset
generator and its implementation generalized for any given UMLS release. We
demonstrate the use of the dataset generator by generating datasets
corresponding to three UMLS releases, 2020AA, 2021AA, and 2021AB. We provide
three UVA baselines using the three existing approaches (LexLM, ConLM, and
RBA). The code, the datasets, and the experiments are publicly available,
reusable, and reproducible with any UMLS release (a no-cost license agreement
is required for downloading the UMLS).

本文提出了一种名为 UVA 的新任务，用于改善 UMLS Metathesaurus 构建流程，并通过逻辑规则和神经网络等多个方面来生成基准数据集和方法。