Long-text repetition loops traced to sampler, not model

A NeurIPS 2026 paper shows that repetition loops in long-form generation stem from top-p sampling degenerating to always pick the top-ranked token, not from model capability. The authors advise checking the sampler before investing in training fixes.

2026-09-28 ~ 2026-09-28 · 3 related posts

• Why LLMs loop on long outputs: top-p ends up picking the #1 token every step — ziv_ravid · 2026-09-28
• Check your sampler before more training: does looping hit agentic coding too? — ziv_ravid · 2026-09-28
• Long-context loops blamed on samplers, not models, in NeurIPS 2026 paper — menhguin · 2026-09-28