跳转至

何时复习：语言模型持续预训练的间隔重复机制

文章背景与核心概要

大语言模型（LLM）在持续预训练过程中常常面临“灾难性遗忘”问题，即模型在获取新知识的同时，会以牺牲先前学到的知识为代价。传统的经验重放方法通常依赖于全局的新旧数据混合以及均匀采样，完全忽略了一个事实：不同的样本被遗忘的速度截然不同。

本文引入了一种受认知科学启发的持续学习框架——间隔重复训练（Spaced Repetition Training, SRT）。SRT 利用 SuperMemo-2 (SM-2) 算法，通过追踪个体复习状态并将困惑度映射为回忆质量，来自适应地调度样本复习。在不改变基础模型、训练目标或优化器的情况下，SRT 成功地改善了稳定性与可塑性之间的权衡，并在语言、视觉和表格领域展现出强大的性能。

When to Review: Spaced Repetition for Continual Pre-Training of Language Models

arXiv: 2608.17530 [cs.AI]
Submitted on: August 18, 2026
Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)

Authors / 作者

  • Alankar Atreya
  • Devesh Batra
  • Yoages Kumar Mantri
  • Geremy Bantug
  • Greig A Cowan
  • Raad Khraishi

Executive Summary / 执行摘要

大语言模型（LLM）的持续预训练往往会遭受“灾难性遗忘”的困扰——即模型在获取新信息的同时，擦除了先前学习到的知识。传统的重放方法通常依赖于全局的新旧数据混合以及统一采样，完全忽略了不同样本被遗忘的速度存在巨大差异这一事实。

本文引入了间隔重复训练（Spaced Repetition Training, SRT），这是一个受认知科学启发的持续学习框架。SRT 利用 SuperMemo-2 (SM-2) 算法，通过跟踪单个复习状态并将困惑度映射到回忆质量，来自适应地调度样本复习。SRT 在不改变基础模型、训练目标或优化器的前提下，成功增强了稳定性与可塑性的权衡，并在语言、视觉和表格领域展示了强健的性能。

Continual pre-training of Large Language Models (LLMs) often suffers from "catastrophic forgetting"—where models acquire new information at the expense of erasing previously learned knowledge. Traditional rehearsal methods typically rely on global old/new data mixtures and uniform sampling, completely overlooking the fact that different examples are forgotten at vastly different rates.

This paper introduces Spaced Repetition Training (SRT), a continual learning framework inspired by cognitive science. SRT utilizes the SuperMemo-2 (SM-2) algorithm to adaptively schedule sample rehearsals by tracking individual review states and mapping perplexity to recall quality. SRT successfully enhances the stability-plasticity trade-off without altering the base model, objective, or optimizer, demonstrating robust performance across language, vision, and tabular domains.

Abstract / 摘要

大语言模型的持续预训练必须在不擦除旧知识的前提下获取新信息。现有的重放方法通常选择全局的新旧混合数据并进行均匀采样，忽略了不同样本被遗忘的速度各有不同。

我们将持续预训练构想为自适应的复习调度：训练循环不仅应决定回放多少历史数据，还应决定在每一步中应该让哪些样本重新出现。我们引入了间隔重复训练（Spaced Repetition Training, SRT），这是一个受认知科学启发的持续学习框架，它使用 SuperMemo-2 (SM-2) 算法来调度样本重放。

SRT 维护每个样本的复习状态，将每个样本的困惑度映射为回忆质量信号，并在保持模型、目标函数和优化器不变的情况下，调度历史样本进行保留、新样本进行巩固。

在时间上分离的 Wikipedia 和代码语料库上，SRT 改善了稳定性与可塑性之间的权衡，在不同模型规模下，恢复了因朴素持续预训练而丢失的 5 到 37 个百分点的旧知识准确率，同时保持或提升了新知识的获取能力。在更大规模下，SRT 保持了广泛的基准性能，而朴素持续预训练和均匀重放则使这些性能大幅下降。视觉和表格数据的实验进一步表明，当与适当的回忆信号配合使用时，这种调度原则可以扩展到语言之外的领域。

Continual pre-training of large language models must acquire new information without erasing old knowledge. Existing replay methods often choose a global old/new mixture and sample uniformly, ignoring that examples differ in how quickly they are forgotten.

We formulate continual pre-training as adaptive review scheduling: the training loop should decide not only how much history to replay, but which examples should return at each step. We introduce Spaced Repetition Training (SRT), a continual learning framework inspired by cognitive science, which schedules sample-rehearsal using the SuperMemo-2 (SM-2) algorithm.

SRT maintains per-example review state, maps per-example perplexity to a recall-quality signal, and schedules historical examples for retention and new examples for consolidation while leaving the model, objective, and optimizer unchanged.

On temporally separated Wikipedia and code corpora, SRT improves the stability-plasticity trade-off, recovering 5 to 37 percentage points of old-knowledge accuracy lost by naive continual pre-training across model scales while preserving or improving new-knowledge acquisition. At larger scale, SRT preserves broad benchmark performance that naive continual pre-training and uniform replay substantially degrade. Experiments with vision and tabular data further suggest that the scheduling principle extends beyond language when paired with an appropriate recall signal.

Key Contributions & Methodology / 核心贡献与方法论

  • 自适应复习调度： 将持续学习构想为一个主动调度问题，而非静态数据混合，并针对每个样本进行优化。
  • SM-2 算法集成： 利用认知科学支持的 SuperMemo-2 算法来管理单个样本的复习状态。
  • 困惑度到质量的映射： 将每个样本的模型困惑度转化为可靠的回忆质量信号。
  • 即插即用兼容性： 无需对底层神经网络架构、训练目标或优化器进行任何修改。
  • 跨域泛化： 证明了其在语言语料库（Wikipedia、代码）、视觉数据和表格数据集上的有效性。

Key Contributions & Methodology

  • Adaptive Review Scheduling: Frames continual learning not as static data mixing, but as an active scheduling problem optimized per sample.
  • SM-2 Algorithm Integration: Leverages the cognitive-science-backed SuperMemo-2 algorithm to manage individual sample review states.
  • Perplexity-to-Quality Mapping: Translates per-example model perplexity into a reliable recall-quality signal.
  • Plug-and-Play Compatibility: Requires zero modifications to the underlying neural network architecture, training objectives, or optimizers.
  • Cross-Domain Generalization: Demonstrates effectiveness across language corpora (Wikipedia, code), vision data, and tabular datasets.