Research map / Continual learning / Continual fine-tuning of LLMs
Improving Language Plasticity via Pretraining with Active Forgetting
Yihong Chen, Kelly Marchisio, Roberta Raileanu, David Ifeoluwa Adelani, Pontus Stenetorp, Sebastian Riedel, Mikel Artetxe
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 |
|---|---|---|---|
| Reset token embeddings periodically during pretraining | RoBERTa-base · XNLI, 5M-token language adaptation | accuracy: 53.3 -> 62.7; +21.2% | Improved |
| Reset token embeddings periodically during pretraining | RoBERTa-base · MLQA, 5M-token language adaptation | F1: 34.3 -> 43.4; +33.8% | Improved |
| Reset token embeddings periodically during pretraining | RoBERTa-base · XQuAD, 5M-token language adaptation | F1: 36.1 -> 49.0; +60.9% | Improved |
| Reset token embeddings periodically during pretraining | RoBERTa-base · English XNLI | accuracy: 86.1 -> 85.1; -1.2% | Hurt |
| Reset token embeddings periodically during pretraining | RoBERTa-base · XNLI Thai after multilingual pretraining | accuracy: 49.4 -> 55.0 | Improved |
| Reset token embeddings periodically during pretraining | RoBERTa-base · XNLI Thai after English monolingual pretraining | accuracy: 35.3 -> 59.7 | Improved |
| Reset token embeddings periodically during pretraining | RoBERTa-base · Arabic NLI with 5M unlabeled tokens | accuracy: 41.2 -> 59.7 | Improved |
| Reset token embeddings periodically during pretraining | RoBERTa-base · XNLI during early language adaptation | accuracy: 37.2 -> 57.8 | Improved |
| Adapt the full model on the task instead of separately adapting embeddings and task body | Standard multilingual RoBERTa-base · XNLI Thai | accuracy: 49.4 -> 60.0 | Improved |
| Use forgetting pretraining with standard labeled adaptation instead of standard pretraining with standard adaptation | RoBERTa-base · Arabic NLI | accuracy: 32.8 -> 34.2 | Improved |
| Set the forgetting interval to K=100 instead of K=1000 (ablation) | Active forgetting pretraining · Pretraining loss | loss: around 11 | Hurt |
| Set the forgetting interval to K=5000 instead of K=1000 (ablation) | Active forgetting pretraining · Pretraining stability | divergence | Hurt |
Closest papers by meaning
- Language Model Memory and Memory Models for Language (NeurIPS 2026 · similarity 0.89)
- Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory (NeurIPS 2026 · similarity 0.89)
- What Will My Model Forget? Forecasting Forgotten Examples in Language Model Refinement (ICML 2024 · similarity 0.88)
- Through the River: Understanding the Benefit of Schedule-Free Methods for Language Model Training (NeurIPS 2025 · similarity 0.88)
- Scaling Laws for Forgetting during Finetuning with Pretraining Data Injection (ICML 2025 · similarity 0.88)
- Spurious Forgetting in Continual Learning of Language Models (ICLR 2025 · similarity 0.88)
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.
