Tag
This paper proposes H-Res, a method to adapt large transformer models by shaping the energy landscape of associative memories without modifying weights or adding prompts, preserving memory capacity and outperforming LoRA.