Installs as a private draft. Your edits and self-improvement runs do not change the published bundle.
Control how far the policy drifts from the reference model during training
Control how far the policy drifts from the reference model during training
Too much drift = reward hacking; too little = no improvement
RL Training Infrastructure > Training Operations
--- name: "Kl Divergence Management" version: "1.0.0" description: "Control how far the policy drifts from the reference model during training" tags: - rl-skills-universe - "rl-training-infrastructure" - "training-operations" trigger: "kl-divergence-management" --- # Kl Divergence Management Control how far the policy drifts from the reference model during training ## Why It Matters Too much drift = reward hacking; too little = no improvement ## Domain **RL Training Infrastructure** > Training Operations