Amazon Nova Uses rDPO for Selective Unlearning
Adapter-scoped safety customisation on a major cloud model family could reshape how APS teams govern sensitive AI workloads on AWS Bedrock.
Key points
- AWS introduced rDPO, a LoRA-adapter technique enabling approved enterprise customers to reduce model over-deflection in selected safety categories.
- The approach separates configurable moderation behaviour from non-configurable protections, potentially relevant for government security, legal, and research workloads.
- All benchmark results are vendor-reported; independent validation of residual risk and governance boundaries is still required before reliance.
Implications for Australian agencies
- Monitor Agencies using AWS Bedrock for sensitive workloads may want to monitor CCMS availability, approved-customer eligibility, and any independent evaluations of residual risk.
- Consider AI governance teams could consider how adapter-scoped safety customisation interacts with existing agency-level AI risk assessments and vendor management obligations under the APS AI Policy.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.
View original source
Copied.
Appeared in:
Weekly digest, 6 July 2026
"Amazon Nova Uses rDPO for Selective Unlearning"
Source: Let's Data Science – AI Governance
Published: 8 July 2026
URL: https://letsdatascience.com/news/amazon-nova-uses-rdpo-for-selective-unlearning-8c7ee1b5
AWS has described Reverse Direct Preference Optimization (rDPO) as the mechanism behind Amazon Nova's Customizable Content Moderation Settings (CCMS), available to approved enterprise customers. The approach uses LoRA adapters to reduce over-deflection in selected responsible AI policy areas while leaving base model weights and non-configurable protections - such as child-safety and privacy controls - intact. AWS reports deflection dropping from 86.51% to 32.77% in one category with utility benchmark losses under two percentage points. The practical signal for government teams is that legitimate sensitive workloads - malware analysis, legal discovery, trust-and-safety review - may become more tractable without broad prompt workarounds, but all results are currently vendor-reported and require independent validation.
Implications for Australian agencies:
- [Monitor] Agencies using AWS Bedrock for sensitive workloads may want to monitor CCMS availability, approved-customer eligibility, and any independent evaluations of residual risk.
- [Consider] AI governance teams could consider how adapter-scoped safety customisation interacts with existing agency-level AI risk assessments and vendor management obligations under the APS AI Policy.
Retrieved from SIMS, 18 July 2026.