Frontier LLMs Fail at Verbatim Copying — and Viewing Text in 2D Fixes It

arXiv·high signal

Haodong Wen, Yiran Zhang and Yingfa Chen (arXiv 2607.16072, cs.CL) show that models capable of advanced multi-step reasoning fail at the far simpler operation of copying text exactly, and that reframing text as a 2D structure rather than a 1D token stream substantially improves fidelity. This is a direct diagnosis of a bug builders hit constantly: agents that mangle long code blocks, config files, or JSON on passthrough. The finding suggests the failure is representational, not a prompting problem — meaning no amount of 'copy this exactly' instruction fully solves it.