AWS demonstrated infrastructure for multi-turn RL training of Amazon Nova model

AWS Demonstrates Infrastructure for Multi-turn RL Training of Amazon Nova Model

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

How the Training Pipeline Works

Following deployment, the result is an event-driven pipeline: training launches automatically at the moment when a user uploads data to an Amazon S3 bucket. As a demonstration task, the model is trained to play the word game Wordle — the authors explicitly call this a "placeholder" that the reader can replace with their own RL task. Wordle is convenient as an educational example precisely because it is a simple game with clear, easily verifiable results on each move — the word is guessed or it isn't — which means it is easy to construct a clear reward signal for the model based on it.

• The infrastructure consists of two phases and launches automatically by event
• The trigger is uploading data to an Amazon S3 bucket
• Training unfolds on Amazon SageMaker HyperPod using Amazon Nova Forge
• The demonstration task is training the model to play Wordle, which serves only as an example and can be replaced with any other RL task

Why Multi-turn RL is Needed

Multi-turn reinforcement learning is training a model through a series of sequential moves or exchanges with feedback, rather than through generation of a single isolated response. This approach is particularly important for agentic AI systems: models that must plan several steps ahead, change strategy during dialogue or play, and learn from the result of an entire sequence of actions rather than a single exchange — for example, customer support assistants, trading or game agents, and tools that invoke external APIs.

Amazon Nova is a family of base models which AWS introduced as part of the Bedrock platform; Amazon Nova Forge in this context serves as a tool for their further training. Amazon SageMaker HyperPod, in turn, is AWS's managed infrastructure for resilient distributed training on large GPU clusters, designed for automatic recovery from individual node failures without losing progress of the entire training run.

Event-driven architecture is important in itself here: instead of manually launching training each time new data appears, the pipeline responds to the very fact of file upload to S3 and starts training automatically. This simplifies the organization of continuous model retraining as new data arrives and reduces the number of manual operations needed for a team to launch the next RL cycle. Automatic recovery of the HyperPod cluster after individual node failures, in turn, reduces the risk that a many-hour or many-day training run will need to be restarted from scratch due to a single hardware failure.

What This Means

The publication shows that AWS continues to expand the toolkit for fine-tuning its own Nova models through reinforcement learning directly on managed infrastructure — which competes with similar RL fine-tuning offers for models from OpenAI, Google, and Anthropic and makes it easier for clients to create their own agentic AI systems on top of Nova. The use of intentionally simplified games like Wordle as a training example also shows that AWS is directing the material not only to research teams but to engineers who need a clear starting example for their own projects.