A new framework called DNAlign uses control-theoretic optimization and null-space projection to align LLMs toward safer outputs while preserving general knowledge. By treating the model as a dynamic system and restricting…
#AI #LLMSafety #Alignment #Cybersecurity
https://arxiv.org/abs/2610.02844
