AI Paper Atlas

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

ICML 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Continual pretraining of language models · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
select replay prompts by attention contribution (STOC) instead of random snippets (ablation)generative data replay, full-parameter · Biography CPT, original knowledge, mix ratio 0.5hFTA: 18.96 -> 55.91Improved
use STOC replay instead of LAMOL special-token promptsPythia-160M, full-parameter · Biography CPT, original knowledge, mix ratio 0.5EM: 5.95 -> 29.84Improved
use STOC replay instead of LAMOLPythia-160M, full-parameter · Biography CPT, new knowledge, mix ratio 0.5EM: 85.33 -> 71.41Hurt
freeze first 6 layers together with STOC (ablation)STOC · Biography CPT, original knowledge, mix ratio 0.5hFTA: 55.91 -> 59.07Improved
use LoRA rank 128 instead of full tuning with STOC (ablation)STOC · Biography CPT, original knowledge, mix ratio 0.5hFTA: 55.91 -> 51.36Hurt
increase CPT data mixing ratio from 0.5 to 0.9 (less replay) (ablation)STOC · Biography CPT, original knowledge, full-parameterhFTA: 55.91 -> 22.03Hurt
add data augmentation (5 biographies per individual) in pre-training (ablation)Pythia-160M · Biography PT test, unseen templatesEM: 2.71 -> 81.27Improved
add data augmentation (5 biographies per individual) in pre-training (ablation)Qwen2.5-0.5B · Biography PT test, unseen templatesEM: 2.94 -> 83.43Improved
use STOC replay instead of Naive CPTPythia-160M · KnowEdit Wiki_Bio, original knowledgesoft token accuracy: 13.22 -> 23.86Improved
use STOC replay instead of Naive CPTQwen2.5-0.5B · KnowEdit ZSRE, continual knowledgesoft token accuracy: 63.28 -> 62.26No clear effect
use STOC replay instead of LAMOL on legal-domain CPT1.7B model · MMLU continual subsets / original subsetsaccuracy: original 39.28 -> 42.49Improved
add EWC-style regularization during CPT (ablation)Pythia and Qwen CPT · Biography CPT, forgetting dynamicshFTA of original knowledgeNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy