FUSE-ing Language Models: Zero-Shot Adapter Discovery for Prompt Optimization Across Tokenizers

TL;DR

This paper introduces FUSE, a cost-effective method for aligning semantic embeddings across different tokenizers and models, enabling improved zero-shot prompt optimization in diverse language model settings.

Contribution

FUSE presents a novel tensor-based approach to approximate adapters that unify embedding spaces across models and tokenizers, facilitating knowledge transfer.

Findings

Effective across vision-language and causal language models

Improves zero-shot prompt optimization performance

Aligns semantic embeddings across different tokenizers

Abstract

The widespread use of large language models has resulted in a multitude of tokenizers and embedding spaces, making knowledge transfer in prompt discovery tasks difficult. In this work, we propose FUSE (Flexible Unification of Semantic Embeddings), an inexpensive approach to approximating an adapter layer that maps from one model's textual embedding space to another, even across different tokenizers. We introduce a third-order tensor-based representation of a model's embedding space that aligns semantic embeddings that have been split apart by different tokenizers, and use this representation to derive an approximation of the gradient of one model's outputs with respect to another model's embedding space. We show the efficacy of our approach via multi-objective optimization over vision-language and causal language models for image captioning and sentiment-based image captioning.