Learning through curiosity: Scholars developed a novel training method called Curious Replay, which programs AI agents to “self-reflect” about things they recently encountered to help them successfully adapt to their changing surroundings. stanford.io/3PN9bEI

I had a great time working on this with Isaac Kauvar and Nick Haber. It is exciting to see the potential of using animal behavior to evaluate AI agents and inspire new approaches!

Goal conditioned RL w/ a surprisingly simple way to do hierarchy ("I can't believe it's hierarchical"): just using a flat (low-level) value function to train a high- and low-level actor turns out to work really well for offline GCRL: seohong.me/projects/hiql/ Hierarchical IQL: 🧵👇