🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

We found a way to do RL *only* with BC policies. The idea is simple: 1. Train a BC policy π(a|s) 2. Train a conditional BC policy π(a|s, z) 3. Amplify(!) the difference between π(a|s, z) and π(a|s) using CFG Here, z can be anything (e.g., goals for goal-conditioned RL). 🧵↓

Cool ideas from Yifei, who generally has great sense on building these self-improving systems

I really liked this work because of the solid science. There are 17 pages of experiments in the appendix… We systematically tried to scale every axis we could think of (data, model size, compute) and over 1000+ trials found only one thing consistently mattered.

Replicated in MLX on MNIST. S+ is an intriguing optimizer that excels at both memorizing the training data and generalizing well. Very intriguing and different from most other optimizers I've tested. Takes some time to get going but typically ends up doing slightly better than

We’re super thrilled to have received the Outstanding Demo Paper Award for MuJoCo Playground at RSS 2025! Huge thanks to everyone who came by our booth and participated, asked questions, and made the demo so much fun! Carlo Sferrazza Qiayuan Liao Arthur Allshire