WEAK-TO-STRONG DECODING FOR LLM ALIGNMENT

Summary from the TLDR AI newsletter

WSD is a novel method where a small aligned model drafts the start of a response, then a large base model continues it. This boosts alignment without harming performance.