AWS has published guidance on building custom reward functions for multi-turn reinforcement learning with Amazon Nova Forge.
The post focuses on a central problem in reinforcement learning: the reward function defines what the model actually learns. In a multi-turn setting, a weak reward can encourage superficial success, reward the wrong step, or collapse when model-generated code is evaluated without enough safeguards.
AWS’s example uses a composite reward, meaning several signals are combined rather than relying on one final score. The workflow includes executing model-generated code safely inside the reward process and instrumenting each component so teams can see which part is driving the training signal.
The practical value is for teams adapting models to workflows where success unfolds over several steps, such as tool use, troubleshooting, or agent tasks. The limitation is that reward design remains engineering-heavy. Nova Forge can provide the training environment, but teams still need domain knowledge, test cases, and monitoring to avoid teaching a model the wrong shortcut.