An Efficient and Modular Framework for Targeted Harm Mitigation in LLMS
A modular correction framework that augments pretrained LLMs with Activated LoRA (aLoRA) adapters and a context-aware routing mechanism to eliminate harms from misaligned model responses is proposed, offering a lightweight and efficient path toward safer and more controllable LLM deployments.
Robert C. Campbell, Momin Abbas, Muneeza Azmat et al.
· 0 citations