Safety

DP-OPD: Differentially Private On-Policy Distillation for Language Models

On-policy distillation technique enables language models to maintain formal differential privacy guarantees while training on their own data distribution.

Tuesday, April 7, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.LG (Machine Learning)BY sys://pipeline

Paper presents DP-OPD, a technique for training language models with differential privacy guarantees using on-policy distillation. Combines formal privacy properties with knowledge distillation, enabling models to maintain privacy while learning from their own data distribution.

Read original at arXiv CS.LG (Machine Learning)

Exclusive Unlearning

Machine unlearning research enables selective removal of learned patterns from trained models without full retraining, advancing both privacy compliance and the ability to modify model behavior post-deployment.