Natively Unlearnable Large Language Models

In three linesNULLs (Natively Unlearnable LLMs) is an architecture that isolates each data source's contributions in distinct parameters (sinks) while maintaining a shared backbone. Tested on ~6M Wikipedia articles, it enables unlearning a specific source at deployment without retraining, while preserving shared knowledge and general language capabilities.

Summary generated by Claude — human-verified