Model WarsJuly 6, 2026via AWS Machine Learning Blog
Teaching models to forget: Selective unlearning with Amazon Nova
Why it matters
Amazon introduces Reverse Direct Preference Optimization (rDPO), a novel unlearning technique that lets models selectively forget content while maintaining performance. This addresses a critical gap in enterprise AI: fine-grained content moderation without degrading overall model quality—directly relevant to regulated industries and custom deployments.
Key signals
- Technique: Reverse Direct Preference Optimization (rDPO)
- Application: Amazon Nova Customizable Content Moderation Settings (CCMS)
- Core benefit: Reduces over-deflection while preserving model quality
- Use case: Selective unlearning for enterprise content policies
- Availability: Customers can apply technique to their own preference optimization experiments
The hook
Amazon Nova just shipped selective unlearning. Here's why every enterprise AI team needs to care about rDPO.
In this post, we introduce Reverse Direct Preference Optimization (rDPO), the novel unlearning technique behind Amazon Nova Customizable Content Moderation Settings (CCMS), and show how it reduces over-deflection while preserving model quality. We also provide pointers for customers who want to apply these preference optimization techniques to their own experiments.