AWS details infrastructure for multi-turn RL training of Amazon Nova

The AWS Machine Learning Blog has published instructions for deploying a two-phase infrastructure for multi-turn reinforcement learning (RL) of the Amazon Nova model, using Amazon Nova Forge on the Amazon SageMaker HyperPod platform.

How the training pipeline works

The deployment produces an event-driven pipeline: training starts automatically as soon as a user uploads data to an Amazon S3 bucket. For the demo task, the guide trains the model to play the word game Wordle. The authors explicitly call it a "placeholder" that readers can swap for their own RL task. Wordle works well as a teaching example precisely because it is a simple game with a clear, easily verifiable outcome on every turn — the word is either guessed or it isn't — which makes it easy to build a clear reward signal for the model.

• The infrastructure has two phases and launches automatically on an event
• The trigger is a data upload to an Amazon S3 bucket
• Training runs on Amazon SageMaker HyperPod using Amazon Nova Forge
• The demo task is teaching the model to play Wordle, which serves only as an example and can be replaced with any other RL task

Why multi-turn RL matters

Multi-turn reinforcement learning trains a model over a series of sequential turns or exchanges with feedback, rather than on a single isolated response. The approach matters most for agentic AI systems: models that need to plan several steps ahead, adjust their strategy over the course of a conversation or game, and learn from the outcome of an entire sequence of actions rather than a single reply. Examples include customer support assistants, trading or gaming agents, and tools that call external APIs.

Amazon Nova is a family of foundation models that AWS introduced as part of the Bedrock platform; in this context, Amazon Nova Forge serves as the tool for fine-tuning them. Amazon SageMaker HyperPod is AWS's managed infrastructure for resilient distributed training on large GPU clusters. It is designed to recover automatically from individual node failures without losing the progress of the entire training run.

The event-driven architecture matters in its own right. Instead of someone manually launching training every time new data shows up, the pipeline reacts to the file upload to S3 and starts training on its own. That makes it easier to set up continuous fine-tuning as new data arrives and cuts the number of manual steps a team needs to kick off each RL cycle. HyperPod's automatic recovery from individual node failures, in turn, lowers the risk that a training run lasting hours or days will have to restart from scratch because of a single hardware fault.

What it means

The guide shows that AWS keeps expanding its tooling for fine-tuning its own Nova models with reinforcement learning directly on managed infrastructure. That puts it in competition with similar RL fine-tuning offerings from OpenAI, Google, and Anthropic, and makes it easier for customers to build their own agentic AI systems on top of Nova. Using a deliberately simple game like Wordle as the training example also signals that AWS is aiming the material not only at research teams but also at engineers who need a clear starting point for their own projects.