AI Paper Atlas

Research map / Continual learning / Continual fine-tuning of LLMs

A Unified Continual Learning Framework with General Parameter-Efficient Tuning

Qiankun Gao, Chen Zhao, Yifan Sun, Teng Xi, Gang Zhang, Bernard Ghanem, Jian Zhang

ICCV 2023 · Continual learning › Continual fine-tuning of LLMs · approach: Low-rank and model-merging adaptation · 135 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
add learning with calibrated adaptation speed (ablation)LAE Adapter10 · ImageNet-RA10: 65.22 -> 72.45Improved
add accumulation and ensemble on top of calibrated learning (ablation)LAE Adapter10 · ImageNet-RA10: 70.30 -> 72.45Improved
use expert ensemble instead of offline PET only (ablation)LAE Adapter10 · ImageNet-RA10: 72.80 -> 72.45 (A10); 78.63 -> 79.07 (average)No clear effect
add EMA accumulation to the ensemble (ablation)LAE Adapter10 without calibrated learning · ImageNet-RA10: 63.36 -> 65.22Improved
add gradient compensation to Prefix (ablation)LAE Prefix20 · ImageNet-RA10: 69.23 -> 70.53Improved
add learnable scaling parameters to Prefix (ablation)LAE Prefix20 · ImageNet-RA10: 69.23 -> 70.38Improved
combine compensation and scale for Prefix (ablation)LAE Prefix20 · ImageNet-RA10: 69.23 -> 72.05Improved
apply LAE framework to baseline PET with Adapter5Adapter5 PET · ImageNet-RA10: 61.63 -> 72.66Improved
apply LAE to baseline with Prefix10Prefix10 PET · CIFAR100A10: 84.49 -> 85.17Improved
increase PET module size (ablation)LAE · ImageNet-RA10: 72.66 -> 72.45 (Adapter5 -> Adapter10)No clear effect
use LAE with Swin-B instead of baselineSwin-B Adapter · ImageNet-RA10: 71.81 -> 73.38Improved
use LAE with ConvNeXt-B instead of baselineConvNeXt-B Adapter · ImageNet-RA10: 76.35 -> 78.38Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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