AI Paper Atlas

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

Continual Low-Rank Adapters for LLM-based Generative Recommender Systems

Hyunsik Yoo, Ting-Wei Li, SeongKu Kang, Zhining Liu, Charlie Xu, Qilin Qi, Hanghang Tong

ICLR 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Low-rank and model-merging adaptation

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 proximal regularizer toward previous LoRA statesingle evolving LoRA · InstrumentsNDCG@5: 0.0127 -> 0.0138Improved
add proximal regularizer toward previous LoRA statesingle evolving LoRA · BooksHit@5: 0.0448 -> 0.0448No clear effect
sum all frozen past adapters (cumulative LoRA) without parameter inheritancesingle evolving LoRA · Amazon InstrumentsNDCG@5: 0.0127 -> 0.0093Hurt
add parameter inheritance to cumulative LoRASumLoRA / cumulative · Instruments natural splitrelative NDCG@5 gain vs single: -26.77% -> 1.57%Improved
use user-disjoint split instead of natural chronological splitcumulative LoRA (all) · Instrumentsrelative NDCG@5 gain vs single: -26.77% -> -8.13%Improved
learnable magnitudes for past adapters (SD-LoRA) vs fixedcumulative LoRA latest+inherit · InstrumentsNDCG@5: 0.0130 -> 0.0128No clear effect
add orthogonality regularizer against previous adapterPESO · regularizer comparisonperformanceHurt
use L2 proximal instead of softmax-KL proximal (ablation)PESO · regularizer comparisonperformanceHurt
use LoRA-output KL or per-rank KL instead of per-module softmax-KL (ablation)PESO · regularizer comparisonperformanceNo clear effect
increase proximal weight lambda from 0 (ablation)PESO · lambda sweepaverage of four metricsMixed
use pretraining learning rate on incremental blocks (lr*=1.0)single evolving LoRA · incremental stagesperformanceHurt
scale learning rate to 0.05-0.1x for incremental blockssingle evolving LoRA · incremental stagesperformanceImproved

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