Ravi Sundaram

A single layer of attention can only compute \( k \)-bit parity with \( k \) heads, revealing fundamental limits on attention mechanisms in neural architectures.

LLMs can automatically translate complex access control rules into plain English, making security policies understandable to non-experts.