A new arXiv paper challenges the common assumption that transformers are already good at copying from their inputs. The authors argue that, despite this intuition, the in-context copying ability of large language models can be improved through explicit modeling.
The proposed approach, Copy-Paste Language Models, is presented as a way to make copying a more deliberate part of how the model operates. The abstract does not include implementation details or evaluation results, so the main claim is that explicit modeling offers a path beyond relying on implicit copying behavior.
Because the source is only the paper's abstract, there are no competing findings to compare. The significance, as stated, is that copying is not just an emergent side effect but something that can be designed for directly.