Aligners: Decoupling LLMs and Alignment

Large Language Models (LLMs) need to be aligned with human expectations toensure their safety and utility in most applications. Alignment is challenging,costly, and needs to be repeated for every LLM and alignment criterion. Wepropose to decouple LLMs and alignment by training aligner models that can beused to align any LLM for a given criteria on an as-needed basis, thus alsoreducing the potential negative impacts of alignment on performance. Our recipefor training the aligner models solely relies on synthetic data generated witha (prompted) LLM and can be easily adjusted for a variety of alignmentcriteria. We illustrate our method by training an "ethical" aligner and verifyits efficacy empirically.