Research map / Continual learning / Continual fine-tuning of LLMs
Towards Understanding Continual Factual Knowledge Acquisition of Language Models: From Theory to Algorithm
Haoyu Wang, yifan shang, Zhongxiang Sun, Weijie Yu, Xiao Zhang, Jun Xu
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 |
|---|---|---|---|
| select replay prompts by attention contribution (STOC) instead of random snippets (ablation) | generative data replay, full-parameter · Biography CPT, original knowledge, mix ratio 0.5 | hFTA: 18.96 -> 55.91 | Improved |
| use STOC replay instead of LAMOL special-token prompts | Pythia-160M, full-parameter · Biography CPT, original knowledge, mix ratio 0.5 | EM: 5.95 -> 29.84 | Improved |
| use STOC replay instead of LAMOL | Pythia-160M, full-parameter · Biography CPT, new knowledge, mix ratio 0.5 | EM: 85.33 -> 71.41 | Hurt |
| freeze first 6 layers together with STOC (ablation) | STOC · Biography CPT, original knowledge, mix ratio 0.5 | hFTA: 55.91 -> 59.07 | Improved |
| use LoRA rank 128 instead of full tuning with STOC (ablation) | STOC · Biography CPT, original knowledge, mix ratio 0.5 | hFTA: 55.91 -> 51.36 | Hurt |
| increase CPT data mixing ratio from 0.5 to 0.9 (less replay) (ablation) | STOC · Biography CPT, original knowledge, full-parameter | hFTA: 55.91 -> 22.03 | Hurt |
| add data augmentation (5 biographies per individual) in pre-training (ablation) | Pythia-160M · Biography PT test, unseen templates | EM: 2.71 -> 81.27 | Improved |
| add data augmentation (5 biographies per individual) in pre-training (ablation) | Qwen2.5-0.5B · Biography PT test, unseen templates | EM: 2.94 -> 83.43 | Improved |
| use STOC replay instead of Naive CPT | Pythia-160M · KnowEdit Wiki_Bio, original knowledge | soft token accuracy: 13.22 -> 23.86 | Improved |
| use STOC replay instead of Naive CPT | Qwen2.5-0.5B · KnowEdit ZSRE, continual knowledge | soft token accuracy: 63.28 -> 62.26 | No clear effect |
| use STOC replay instead of LAMOL on legal-domain CPT | 1.7B model · MMLU continual subsets / original subsets | accuracy: original 39.28 -> 42.49 | Improved |
| add EWC-style regularization during CPT (ablation) | Pythia and Qwen CPT · Biography CPT, forgetting dynamics | hFTA of original knowledge | No clear effect |
Closest papers by meaning
- How Do Large Language Models Acquire Factual Knowledge During Pretraining? (NeurIPS 2024 · similarity 0.91)
- Spurious Forgetting in Continual Learning of Language Models (ICLR 2025 · similarity 0.89)
- Train-Attention: Meta-Learning Where to Focus in Continual Knowledge Learning (NeurIPS 2024 · similarity 0.89)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.88)
- Scalable Language Model with Generalized Continual Learning (ICLR 2024 · similarity 0.88)
- A Unified and General Framework for Continual Learning (ICLR 2024 · 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.
