Agentic RL

Gradient descent trains a model on data somebody wrote; reinforcement learning trains it on its own attempts. Sample a group of answers, score them, and let the group's average — not a second network — decide which ones deserve more probability.

1 exhibit Computed teaching policy Offline, single file

StudioStart hereComputed
RL Studio

Sample a group of responses, shape the reward yourself, and train a small policy with GRPO: group-relative advantages, optional PPO clipping, and an exact KL leash to the reference model — plus a length trap that shows proxy reward and true quality parting ways.
Single screen · DOM + MathML→

Agentic RL

RL Studio