AI Insight
Researchers developed Spaced Repetition Training (SRT), a method that helps large language models learn new information while retaining old knowledge during continual pre-training. Inspired by cognitive science principles used in human learning, SRT uses the SuperMemo-2 algorithm to intelligently schedule which training examples should be reviewed and when, based on per-example difficulty rather than uniform sampling. Testing on Wikipedia and code datasets showed SRT recovered 5-37 percentage points of accuracy on old knowledge that was lost during standard continual training, while maintaining or improving performance on new material across different model sizes.
Why it matters
This approach addresses a critical challenge in keeping AI language models up-to-date without expensive full retraining from scratch. The technique could enable more efficient and cost-effective updates to deployed language models, and preliminary results suggest the scheduling principle may extend to other data types beyond text.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Abstract: 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.
Source: When to Review: Spaced Repetition for Continual Pre-Training of Language Models