Josh Engels(@JoshAEngels):Cool new post with Lily tracking where LLM values come from. My takeaways: SFT is (again) a big deal for model behavior, and there’s lots of important and low hanging post-training science still to do!

Josh Engels

@JoshAEngels

Member of technical staff @ METR Previously Google DeepMind, MIT. Opinions my own.

Joined December 2021

142 Following 5.9K Followers

Josh Engels@JoshAEngels

2026.07.09 23:28

Cool new post with Lily tracking where LLM values come from. My takeaways: SFT is (again) a big deal for model behavior, and there’s lots of important and low hanging post-training science still to do!

lily sun@lilysun004

2026.07.09 20:56

New blog post! LLMs exhibit value preferences (e.g. intellectual integrity🤓, warmth🥰) which affect their responses to subjective user queries. These change during post-training, sometimes unexpectedly. We ask: can we predict these value changes from just training data? Maybe!🧵