What happened?

On August 14, 2026, Amazon Web Services (AWS) published a technical blog post explaining custom reward function design for multi-turn reinforcement learning (RL) on the Amazon Nova Forge platform. The post serves as a guide for developers looking to customize Amazon Nova models using reinforcement fine-tuning (RFT).

Nova Forge's Bring Your Own Orchestration (BYOO) feature allows developers to run reward logic in their own environment, such as on Amazon Elastic Container Service (Amazon ECS). The platform also offers a serverless multi-turn RL option, now generally available, for teams that don't want to manage this infrastructure.

Why does it matter?

In multi-turn tasks -agent-based processes involving steps like tool calling, code execution, or error recovery- designing the right reward function directly determines what the model learns. According to AWS, a slightly flawed reward can teach the model the wrong behavior even if training curves look healthy.

For this reason, AWS Lambda-based reward functions used in single-turn tasks fall short for multi-turn tasks, since long conversations and scoring operations can exceed the 15-minute Lambda invocation limit. Nova Forge overcomes this with BYOO by offloading each rollout to the customer environment and collecting completed episodes back for training.

How does the reward function work?

According to information shared by AWS, the system uses the Group Relative Policy Optimization (GRPO) algorithm. In this method, multiple completions (rollouts) are sampled from the model for each conversation and ranked by the reward function; the highest-scoring completions are used to update the model weights.

  • Outcome rewards (episode-level): Measure whether the final output meets the goal, such as whether unit tests pass.
  • Behavioral rewards (turn-level): Measure intermediate behaviors, such as the model asking before acting or calling the correct tool.
  • Penalties: Discourage failure modes such as guessing, repeating, or stalling.

AWS states that using these three components together ensures the model learns both the correct behavior and the correct outcome, otherwise one component could overshadow the other.

What's next?

The blog post is presented as the second installment in AWS's series on this topic; the first part covered Amazon SageMaker HyperPod and Nova Forge infrastructure, along with training configuration. AWS announced it has shared code examples in the aws-samples/sample-nova-multi-turn-rl-infra repository and will cover examples of flawed reward design from an actual training run in a follow-up post.

The trap inside reward design

AWS's warning that a slightly wrong reward teaches the wrong behaviour while the curves look healthy is the most important line in the post. The second component of the three-part structure it describes — turn-level behavioural rewards — is exactly where reward hacking lives.

Reward the model for asking a question before acting and the model learns to ask questions, not to understand the task. It will ask when it is not needed too, because the reward attaches to the behaviour rather than to its appropriateness. Unbalanced by the outcome reward, intermediate incentives of this kind can improve the measured metric while degrading real performance.

The second thing BYOO introduces is about responsibility rather than engineering: rollouts run inside the customer's environment. Beyond the compute cost, that means model-generated code executes on your infrastructure — isolation, resource limits and network access become part of the training configuration.