On StateBench’s hardest level, a flexible Gated DeltaNet handled sequences ~10× longer than those seen in training.
摘要
在 StateBench 最难级别上，灵活的 Gated DeltaNet 处理的序列长度约为训练中所见序列的 10 倍。 我们的新证明表明，即使特征值非负的 Gated DeltaNet，也能在带舍入的固定精度算术下解决某些周期性任务。
摘要由模型根据原文撰写，以原文为准。
正文 · 原文
On StateBench’s hardest level, a flexible Gated DeltaNet handled sequences ~10× longer than those seen in training.
Our new proof shows that even Gated DeltaNet with non-negative eigenvalues can solve some periodic tasks under fixed-precision arithmetic with rounding.
来源：Ai2 · x.com
引用原文
allen_ai. On StateBench’s hardest level, a flexible Gated DeltaNet handled sequences ~10× longer than those seen in training.[EB/OL]. (2026-11-04)[访问日期]. https://x.com/allen_ai/status/2117670301332910258.