The paper itself was called kind of a cheeky title, your language model is secretly a reward model. And that is hinting at this idea that they have found a pretty important mathematical property. They can find the optimal optimization of language model based on these preference of different completions. Instead of estimating a new reward model and training via trial and error, like you said, if you have a preference data, you can just directly do a one step optimization.

Get the Snipd
podcast app

Unlock the knowledge in podcasts with the podcast player of the future.
App store bannerPlay store banner

AI-powered
podcast player

Listen to all your favourite podcasts with AI-powered features

Discover
highlights

Listen to the best highlights from the podcasts you love and dive into the full episode

Save any
moment

Hear something you like? Tap your headphones to save it with AI-generated key takeaways

Share
& Export

Send highlights to Twitter, WhatsApp or export them to Notion, Readwise & more

AI-powered
podcast player

Listen to all your favourite podcasts with AI-powered features

Discover
highlights

Listen to the best highlights from the podcasts you love and dive into the full episode