Combining methods improves memory in language models over time
Continual Learning Mechanisms Compose for Long-Horizon Memorization
Machine Learning
Summary
Remembering lots of tasks over a long time is hard for language models because they tend to forget earlier information as they learn new things. The authors study this problem by training models on many tasks one after another without reminders, to see how well they keep what they learned. They find that using multiple techniques together, each protecting different parts of the model's memory, helps a lot more than using any single method alone. Their best approach shows a big improvement in how much the model remembers after lots of tasks. This suggests that carefully combining approaches can help AI systems remember knowledge more reliably as they learn continuously.
language modelcontinual learningcatastrophic forgettingfine-tuningmemorizationlow-rank adaptationtask learningmodel retentiondata anchorssuccessive halving
Authors
Zheyuan Zhang, Alvin Zhang, Daniel Khashabi, Tianmin Shu
Abstract
Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.