Long-context loops blamed on samplers, not models, in NeurIPS 2026 paper

menhguin · x · 2026-09-28

Researchers announce their long-context sampling paper is accepted to NeurIPS 2026. Ask an open model for a really long story and the second half usually turns into loops well before context runs out — and a big part of the blame lies with the sampler, not the model or data. With top-p, by the end the model picks its #1 choice every step, which is the loop; adaptive samplers avoid this. Authors suggest checking your sampler before spending on more training, and ask whether this holds for long agentic coding runs.

Related event: Long-text repetition loops traced to sampler, not model(3 posts)→

Original post →

More from Models

Models channel →