Can you control a large language model without breaking its brain? Zhejiang University and Alibaba Group researchers just showed how. They unify all model control methods (fine-tuning, LoRA, activation edits) into one single framework, separating effects into "preference"

→ View original post on X — @jiqizhixin