Amazon's New AI Training Makes Small Models Smarter for Complex Tasks

Amazon SageMaker AI has introduced a new training method called multi-turn reinforcement learning (MTRL) that teaches AI models to perform tasks requiring a series of steps, such as processing an expense report or triaging a support ticket. Unlike traditional training that focuses on single answers, MTRL gives feedback only on whether the entire sequence succeeded, allowing the model to learn from trial and error. In a demonstration using the game Wordle, a small AI model improved its win rate from 61% to 84% on unseen words, all for about $339 in training costs.

This matters because most AI models are trained on general data and may not excel at your specific business processes. With MTRL, you can take a low-cost model and train it on your own multi-step workflows, making it more effective without breaking the bank. The service is serverless, so you don't need to manage infrastructure, and you pay only for the tokens used during training. Your data and trained model stay private in your AWS account, ensuring security and compliance.

The catch is that MTRL requires careful design of the reward system—the signal that tells the AI whether it succeeded. If the reward is poorly designed, the model might learn the wrong behavior. Also, while the Wordle example is simple, real-world tasks may be more complex and require more training data and tuning. Additionally, MTRL currently supports only a few base models, so you may need to adapt your approach based on what's available.