跳到正文
arXiv cs.AI· Khizar Qureshi·· 2 小时前AI 评分29

无重放的循环状态迁移信息极限

Information Limits of Replay-Free Recurrent State Migration

AI 导读

arXiv 论文刻画了替换循环模型前必须保留的最小因果记忆:对时不变仿射线性族,k 阶局部迁移所需记忆等于提升系数系统的可达维数减去锚定状态信息。可控的 d 状态标量输入模型最多需要 kd 个额外实坐标,单个 rank-one 方向即可要求全部 kd。实验显示八维历史坐标在五个拟合种子上达到归一化输出误差 10⁻⁴,分裂极点将最小成功宽度提升至 64。

正文 · 原文

View PDF HTML (experimental)

Abstract:When a recurrent model is replaced, its old state may not determine what the new model would compute from the same history. We characterize the information that must be retained before the replacement is known. For a time-invariant affine linear family, the minimum continuous causal memory for order-$k$ local migration equals the reachable dimension of a lifted coefficient system minus the information in the anchor state. A controllable $d$-state model with scalar input needs at most $kd$ additional real coordinates, even when arbitrary matrix replacement directions are revealed only after streaming. A single rank-one direction can already require all $kd$. Conditional singular values and auxiliary-bit bounds distinguish this local information law from recovery at a fixed tolerance. We also give invariant-subspace and future-output constructions. Controlled experiments compare the resulting representations. In a small language model, equal-memory cubic interpolation improves mean state fidelity over third-order coefficients but provides no meaningful predictive gain. An analytic history basis preserves useful adaptation more economically in a coupled identification task. With repeated slow poles, eight history coordinates meet normalized output error $10^{-4}$ across five fitted seeds despite substantial full-state error. Splitting the poles raises the smallest tested successful width to 64.
Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Cite as: arXiv:2611.19345 [cs.AI]
  (or arXiv:2611.19345v2 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2611.19345

arXiv-issued DOI via DataCite

Submission history

From: Khizar Qureshi [view email]
[v1] Sat, 14 Nov 2026 11:34:00 UTC (121 KB)
[v2] Tue, 17 Nov 2026 05:27:32 UTC (125 KB)

来源:arXiv cs.AI · arxiv.org