DYNA cuts LLM forgetting by 7% with temporal knowledge graphs
New framework lets frozen models learn continuously without retraining, outperforming RAG.
Large Language Models (LLMs) suffer from catastrophic forgetting when updated with new knowledge, often requiring costly retraining. A new paper on arXiv introduces DYNA (Dynamic Episodic Memory Networks), a lightweight framework that sidesteps this by attaching an external temporal knowledge graph to a frozen LLM. The graph stores events as nodes and temporal relations as directed, timestamped edges, acting as an updatable episodic memory. At inference time, DYNA retrieves relevant subgraphs using random walks and centrality measures, then augments the LLM's response with context from the graph.
Tested on three temporal recall tasks, DYNA reduces catastrophic forgetting by ~7% compared to fine-tuning, and improves temporal ordering by ~5% over standard retrieval-augmented generation (RAG). The researchers also found that graph clustering coefficients correlate with retrieval performance, suggesting graph structure matters. DYNA requires no retraining of the base LLM, making it practical for continuous learning scenarios like news updates or dynamic databases. Contributions include episodic memory as a temporal KG, retraining-free augmentation, and graph properties as predictors of retrieval success.
- DYNA uses a temporal knowledge graph as external memory, with timestamped edges between event nodes.
- Achieves ~7% reduction in catastrophic forgetting and ~5% better temporal ordering vs. RAG and fine-tuning.
- Graph clustering coefficients predict retrieval accuracy, linking graph topology to memory quality.
Why It Matters
Enables LLMs to stay current with new events without costly retraining, critical for real-time applications.