AI Paper Atlas

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

Merge before Forget: A Single LoRA Continual Learning via Continual Merging

Fuli Qiao, Mehrdad Mahdavi

ICLR 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Low-rank and model-merging adaptation · 6 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
initialize new-task LoRA from last fine-tuned point instead of random/zero (ablation)SLAO · Standard CL Benchmarkaverage accuracy (%): 65.7 -> 80.4Improved
initialize from last fine-tuned point instead of last merged point (ablation)SLAO · SuperNIaverage accuracy (%): 34.0 -> 37.2Improved
initialize from last fine-tuned point instead of last merged point (ablation)SLAO · Standard CL Benchmarkaverage accuracy (%): 80.3 -> 80.4No clear effect
merge only B with time-aware scaling and take A from fine-tuning, instead of fine-tune BA and merge B only (ablation)SLAO · SuperNIaverage accuracy (%): 33.8 -> 37.2Improved
merge B instead of merging A (ablation)SLAO · SuperNIaverage accuracy (%): 25.9 -> 33.8Improved
freeze A and merge B instead of freeze B merge A (ablation)LoRA continual merging · Large Number of Tasksaverage accuracy (%): 69.7 -> 72.9Improved
use SLAO instead of O-LoRALlama-2-7B-chat · SuperNIaverage accuracy (%): 25.9 -> 37.2Improved
use SLAO instead of SAPT-LoRA (which uses generated data)Llama-2-7B-chat · SuperNIaverage accuracy (%): 50.9 -> 37.2Hurt
use SLAO instead of MagMaxLlama-2-7B-chat · Standard CL Benchmarkaverage accuracy (%): 80.3 -> 80.4No clear effect
scale base model from Llama-2-7B-chat to Llama-2-13B-chat (ablation)SLAO · SuperNIaverage accuracy (%): 37.2 -> 42.3Improved
use time-varying scaling 1/sqrt(i) instead of fixed merge factors 0.1/0.5/0.9 (ablation)SLAO · three CL benchmarksaverage accuracyImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 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