Custom Reward Functions for Multi-Turn Reinforcement Learning with Amazon Nova Forge
In brief
- Amazon's Nova Forge introduces a new approach to multi-turn reinforcement learning (RL), focusing on custom reward functions that guide AI models to learn correctly over extended interactions.
- Traditional RL can fail silently even when training looks good because subtle errors in reward design lead the model astray.
- Nova Forge addresses this by allowing users to define precise reward criteria, ensuring models behave as intended across multiple steps-like calling tools or recovering from mistakes.
- The key innovation is Reinforcement Fine-Tuning (RFT), which differs from standard methods by learning from evaluation signals on the model's outputs instead of relying on curated examples.
- This makes it easier for developers to train models iteratively through feedback, rather than needing detailed reasoning paths.
- Multi-turn RFT optimizes cumulative rewards across sequences of actions, leading to better generalization and performance in diverse tasks.
- For teams without the infrastructure to manage environments, Nova Forge offers a serverless multi-turn RL option.
- This feature is now generally available, providing flexibility for different user needs.
- Developers should focus on defining clear reward functions while Nova Forge handles coordination during training.
- The approach emphasizes safety by executing model-generated code securely within reward calculations and monitoring each component's contribution to ensure reliable learning outcomes.
- Looking ahead, this development could significantly improve the reliability of AI systems in complex, multi-step tasks-where subtle design choices can make or break performance.
Terms in this brief
- Reinforcement Fine-Tuning
- A method where AI models learn by adjusting their behavior based on feedback from evaluations of their outputs. Unlike traditional methods that rely on pre-curated examples, RFT allows for iterative training through user feedback, making it easier to refine model performance without needing extensive predefined reasoning paths.
Read full story at AWS ML Blog →
More briefs
New Gemini 3.7 Flash Model Released
Google introduced Gemini 3.7 Flash, a more intelligent model for coding and agents. This new model comes just three weeks after the previous version. Gemini 3.7 Flash has many improvements, including better performance in coding tasks and web development. It can generate more functional layouts and feature-complete apps in fewer prompts. The model also shows high design adherence and parity based on a reference input. Gemini 3.7 Flash will be available at half the original price of the previous model. It delivers substantial improvements across software engineering and knowledge work. The future of coding and agents will likely be shaped by this new model.
Drones Compete in DARPA Lift Challenge
Drones recently competed in the DARPA Lift Challenge. The challenge tested the drones' ability to lift heavy objects. The challenge included many unique drones. Some drones were able to lift objects that weighed over 50 pounds. This matters because it could help drones be used in search and rescue missions. New drone technologies are being developed every day and will likely be used in many industries soon.
Anthropic Explains Chatbot Watermarking
Anthropic published a blog post to answer questions about watermarking its chatbot Claude. The company will add a hidden pattern to the text generated by Claude. The watermarking is required by the EU AI Act’s Transparency Code. This code makes AI companies identify AI-generated content. Anthropic said the watermark will not change the quality of Claude’s output. The company plans to release a watermark detection API. Someone can hide the watermark by completely rewriting the text. The future of AI-generated content will be closely watched as more companies comply with the new regulations.
Google Allows Users to Remove Visible Watermark from AI Generations
Google will now allow users to remove a visible watermark from its AI generations, including images, videos, and songs. This change will affect models like Nano Banana, Omni, and Lyria. The visible watermark was making content less useful for professional and creative work. But Google still wants to identify AI-generated content. The company will keep using invisible SynthID watermarks and metadata for transparency. Over 100 million users will have access to this new feature. Users will be able to turn the visible watermark on or off in the coming days. They can go to Settings to make this change. Google will continue to update its AI features.
AI Satellites to Track Wildfires
Researchers are testing a system that uses artificial intelligence to detect wildfires and reposition satellites for more frequent data. This system could help firefighters detect and fight wildfires sooner. The system could capture more information about wildfires, especially in remote areas. Satellites can detect fires and move to get better views. Current systems use fixed orbits, but this new system allows satellites to reposition themselves. Firefighters will get critical information faster.