Twist Decoding: Diverse Generators Guide Each Other

TL;DR

Twist decoding is a novel inference algorithm that enables the effective combination of diverse language models without shared vocabularies, leading to improved generation quality in tasks like translation and summarization.

Contribution

We introduce Twist decoding, a flexible inference method that leverages multiple diverse models simultaneously without requiring shared vocabularies or tokenization schemes.

Findings

Outperforms individual models across tasks

Surpasses reranking heuristics in quality

Effective with domain-specific and general models

Abstract

Many language generation models are now available for a wide range of generation tasks, including machine translation and summarization. Combining such diverse models may lead to further progress, but ensembling generation models is challenging during inference: conventional ensembling methods (e.g., shallow fusion) require that the models share vocabulary/tokenization schemes. We introduce Twist decoding, a simple and general text generation algorithm that benefits from diverse models at inference time. Our method does not assume the vocabulary, tokenization or even generation order is shared. Our extensive evaluations on machine translation and scientific paper summarization demonstrate that Twist decoding substantially outperforms each model decoded in isolation over various scenarios, including cases where domain-specific and general-purpose models are both available. Twist decoding also consistently outperforms the popular reranking heuristic where output candidates from one model are rescored by another. We hope that our work will encourage researchers and practitioners to examine generation models collectively, not just independently, and to seek out models with complementary strengths to the currently available models. Our code is available at https://github.com/jungokasai/twist_decoding.