Learning Length-Extrapolatable Recurrent Models

2026-09-08

AI Summary

研究解决了长序列建模中训练范围外性能下降的问题，通过提出Credit Stabilization through Time (CST)方法来稳定状态信用信号，从而提高模型在超出训练长度上的表现。

Abstract

Recurrent models provide a natural path to long-context modeling, yet models trained with backpropagation through time (BPTT) often fail beyond their training horizon. Classical analyses emphasize gradients that vanish or explode along temporal paths. However, dense per-token losses can still train a shared recurrent rule despite severe decay, showing that decay alone does not determine whether learning fails. We instead study state credit: the signal through which future losses reach earlier recurrent states before contributing to parameter updates. Accordingly, we intervene directly on state credit and propose Credit Stabilization through Time (CST). During backward propagation, CST locally rescales the state-credit signal to stabilize its norm without rotating the component being corrected, while leaving the forward computation unchanged. Because controlled synthetic tasks and real data exhibit different credit dynamics, we specialize CST to each regime. In both settings, CST improves performance beyond the training horizon, with gains observed at up to 128x the training length.

Problem

• Recurrent Models
• Backpropagation Through Time (BPTT)
• Training Horizon

Innovation

• Credit Stabilization through Time
• state credit
• long-context modeling