Amazon Nova Forge lets you build custom AI rewards

Amazon Nova Forge lets you design custom reward functions for multi-turn reinforcement learning with Nova models. With Bring Your Own Orchestration (BYOO), Nova Forge handles rollouts, message passing, and conversation state while you focus on defining what a good outcome looks like. A serverless multi-turn RL option is also generally available for teams that prefer not to manage that environment. Instead of relying on curated examples like supervised fine-tuning, reinforcement fine-tuning learns from evaluation signals on the model's own outputs, optimizing cumulative reward across entire agent trajectories. According to the article, RL improves out-of-distribution generalization across all task variants while supervised fine-tuning degrades it.