Composing learning mechanisms improves model memory for long tasks
Continual Learning Mechanisms Compose for Long-Horizon Memorization
Machine Learning
Summary
Language models often struggle to remember information from many tasks learned one after another because new learning overwrites old knowledge. The authors studied this problem with 100 tasks learned sequentially and found that combining different techniques that each protect memory in unique ways greatly helps models remember better over time. Their best combination improved memory retention by 28 times compared to simple fine-tuning. This shows that mixing different memory-preserving methods can tackle long-term forgetting more effectively than any single method alone.
What this means in practice
- •For machine learning engineers: Implement training methods that combine multiple memory-preserving mechanisms to improve long-term retention of sequential tasks in language models.
- •For ai product developers: Build language-based applications that update continuously without losing performance on previously learned knowledge using combined continual learning techniques.$Commercial implications: Improves practical deployment of language models for products requiring incremental updates with long-term memory.
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.