Learning to Decode Collaboratively with Multiple Language Models

arXiv:2403.03870v1 Announce Type: cross
Abstract: We propose a method to teach multiple large language models (LLM) to collaborate by interleaving their generations at the token level. We model the decision of which LLM generates the next token as a latent variable. By optimizing the marginal likelihood of a training set under our latent variable model, the base LLM automatically learns when to generate itself and when to call on one of the ``assistant'' language models to generate, all without direct …

arxiv cs.cl cs.lg decode language language models multiple type