compartmentalization

David Chisnall presents CHERI, a hardware capability architecture enabling memory safety and fine-grained compartmentalization through ISA extensions such as ARM Morello, RISC-V, and CHERIoT microcontrollers. The talk emphasizes the importance of practical programming models for secure isolation.

This paper investigates compartmentalization in LLMs, where models fail to share statistical strength across distinct representations of the same concept, leading to reduced sample efficiency and model capacity. The authors demonstrate this phenomenon in multilingual and multi-format settings and show that synthetic parallel data does not fully resolve it.