Why deleting data can fail in agent networks
The study shows deleted data can still echo through later training rounds in self-improving federated agent networks.
Why it matters
It raises the cost and complexity of data deletion in systems that keep learning after launch.
Key facts
- The paper argues that unlearning gets harder when agents keep training after deployment.
- If a data owner asks for deletion, the original data may already have shaped later trajectories that are kept.
- The authors say that influence can survive retraining, grow with more forget-shaped data, and be traced through deployment, collection, and aggregation records.


