← sources

„Learning" without gradients? - LLMs are capable of editing Robot Policies directly in Weight space

source: reddit (reddit:reinforcementlearning)published 2026-11-11 12:54ingested 2026-11-12 02:41

retained text

Teaching your Robot new skills has never been easier:    
Get some data, do behavior cloning, deploy.    
At least this is what many do and it often fails. However, you can just give this policy to a LLM: it will fix the policy for you!    
Full read:    
https://jstranghoener.github.io/blog/robotic-weight-space-programming

document fetched 2026-11-12 00:41 · canonical https://reddit.com/r/reinforcementlearning/comments/1xth0ik/learning_without_gradients_llms_are_capable_of