Paper says frontier models still fail at exact copying as sequences get longer
eyishazyer · x · 2026-07-27
A reply summarizing a paper that argues frontier language models can still fail at exact copying even when they solve advanced reasoning tasks.
- Copying gets less reliable as sequences get longer or more repetitive.
- The issue is partly architectural: with standard 1D RoPE, generating output token yk depends on retrieving xk from a relative offset that shifts with input length.
- The post frames this as a reminder that architecture choices can strongly affect seemingly simple behaviors like verbatim repetition.
Related event: 1D-RoPE Flaw Causes Frontier LLMs to Fail at Long String Copying(2 posts)→
More from Research
- A closed-loop system ran 633 experiments to find a two-metal N2O catalyst — bravo_abad · 2026-07-28
- A deep dive into Stable Diffusion’s safety filter and how red-teaming probes it — vboykis · 2026-07-28
- Lilian Weng’s long guide breaks down diffusion models from noise to samples — vboykis · 2026-07-28
- Nature reports a programmable CRISPR enzyme that shreds cancer-cell DNA — Dr_Singularity · 2026-07-28
- NeurIPS 2026 authors flag a rebuttal workflow bug as reviewers lose access to “Official comments” — howtorewriteaname · 2026-07-28
- Paper finds RLVR improves sampling efficiency but not new reasoning patterns — jxmnop · 2026-07-28