AWS details custom reward functions for multi-turn reinforcement learning in Amazon Nova Forge

Amazon Nova Forge now supports custom reward functions for multi-turn reinforcement learning, a capability aimed at shaping model behavior through iterative feedback rather than curated examples. The approach centers on defining what constitutes a good outcome across multiple interaction turns, which the company argues is one of the hardest parts of customizing Amazon Nova models.

To execute reward logic, developers can use Amazon Nova Forge’s Bring Your Own Orchestration (BYOO) capability, which runs the reward function in a user-controlled environment. This allows teams to focus on defining reward criteria while Nova Forge manages rollouts, message passing, and conversation state across turns. Alternatively, a serverless multi-turn reinforcement learning option is available and described as generally available for teams that prefer not to manage the orchestration environment.

The post warns that subtly incorrect reward functions can quietly teach models the wrong behaviors even when training metrics appear healthy. It recommends designing composite rewards and instrumenting each component to detect pitfalls that could lead to reward collapse, a scenario where the reward signal no longer reflects true performance.

Reinforcement fine-tuning (RFT) is highlighted as a method that learns from evaluation signals on the model’s own outputs, contrasting with supervised fine-tuning’s reliance on annotated reasoning paths. Multi-turn RFT extends this approach to agentic workflows involving tool calls, code execution, or other sequential actions.