Research map / Continual learning / Online and replay-based
Episodic Memory in Lifelong Language Learning
Cyprien de Masson d'Autume, Sebastian Ruder, Lingpeng Kong, Dani Yogatama
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine sparse experience replay with nearest-neighbor local adaptation | BERT-base lifelong learning · Text classification | accuracy: 68.8 -> 70.6 | Improved |
| Combine sparse experience replay with nearest-neighbor local adaptation | BERT-base lifelong learning · Question answering | F1: 60.3 -> 62.4 | Improved |
| Add sparse replay to the encoder-decoder learner | BERT-base lifelong learning · Text classification | accuracy: 18.4 -> 57.8 | Improved |
| Add sparse replay to the encoder-decoder learner | BERT-base lifelong learning · Question answering | F1: 53.1 -> 57.9 | Improved |
| Use nearest neighbors rather than random memory examples for local adaptation (ablation) | MbPA-ARC · Text classification | accuracy: 58.2 -> 70.6 | Improved |
| Use nearest neighbors rather than random memory examples for local adaptation (ablation) | MbPA-ARC · Question answering | F1: 59.7 -> 62.4 | Improved |
| Reduce memory capacity from 100% to 10% (ablation) | MbPA-ARC · Text classification | accuracy: 70.6 -> 67.6 | Hurt |
| Reduce memory capacity from 100% to 10% (ablation) | MbPA-ARC · Question answering | F1: 62.0 -> 61.5 | Hurt |
| Reduce memory capacity from 100% to 50% (ablation) | MbPA-ARC · Text classification | accuracy: 70.6 -> 70.3 | Hurt |
| Reduce memory capacity from 100% to 50% (ablation) | MbPA-ARC · Question answering | F1: 62.0 -> 61.6 | Hurt |
| Increase the number of local-adaptation neighbors from K=8 to K=128 (ablation) | MbPA-ARC · Text classification | accuracy: 68.4 -> 71.6 | Improved |
| Increase the number of local-adaptation neighbors from K=8 to K=32 (ablation) | MbPA-ARC · Question answering | F1: 60.2 -> 62.0 | Improved |
Closest papers by meaning
- Memorization Sinks: Isolating Memorization during LLM Training (ICML 2025 · similarity 0.87)
- Scalable Language Model with Generalized Continual Learning (ICLR 2024 · similarity 0.87)
- CMT: A Memory Compression Method for Continual Knowledge Learning of Large Language Models (AAAI 2025 · similarity 0.87)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.87)
- Online Adaptation of Language Models with a Memory of Amortized Contexts (NeurIPS 2024 · similarity 0.87)
- Human-inspired Episodic Memory for Infinite Context LLMs (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
