Cosmin goes over Score Centering Stabilizes Off-policy Reinforcement Learning
Details
Cosmin goes over Score Centering Stabilizes Off-policy Reinforcement Learning
This Tuesday Cosmin will cover "Score Centering Stabilizes Off-policy Reinforcement Learning". This method is used in the recently thinking machines rl platform update.
Resources
Paper https://arxiv.org/abs/2609.20807
Author thread about the relation to the thinker StableReinforce objective https://x.com/mrtnm/status/2114145574987702525
Other papers and blogs
http://joschu.net/blog/kl-approx.html
https://arxiv.org/abs/2610.22263
https://huggingface.co/blog/NormalUhr/sfc-kl
https://huggingface.co/blog/NormalUhr/sfc
Related topics
Artificial Intelligence
Deep Learning
Machine Intelligence
Machine Learning
Neural Networks