AI Paper Atlas

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

DeLo: Dual Decomposed Low-Rank Experts Collaboration for Continual Missing Modality Learning

Xiwei Liu, Yulong Li, Feilong Tang, Imran Razzak

AAAI 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Mixture-of-experts for multimodal continual tuning · 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
add Cross-Modal Guided Routing (ablation)DeLo · UPMC-Food101-CMML (eta=70%)AP: 48.84 -> 67.42Improved
add Modality-Specific Factor Pool (separate vision and text pools) (ablation)DeLo · UPMC-Food101-CMML (eta=70%)AP: 56.52 -> 67.42Improved
use dynamic decomposed LoRA experts instead of static LoRA (ablation)DeLo · UPMC-Food101-CMML (eta=70%)AP: 60.39 -> 67.42Improved
increase selected rank from 2 to 4 (ablation)DeLo (E=16) · UPMC-Food101-CMML (eta=70%)AP: 63.78 -> 67.42Improved
increase selected rank from 4 to 8 (ablation)DeLo (E=16) · UPMC-Food101-CMML (eta=70%)AP: 67.42 -> 67.11No clear effect
increase factor pool from E=16 to E=32 at r=4 (ablation)DeLo · UPMC-Food101-CMML (eta=70%)AP: 67.42 -> 68.03Improved
add consistency loss to classification loss (ablation)DeLo · UPMC-Food101-CMML (eta=70%)AP: 61.95 -> 65.15Improved
add alignment loss to classification loss (ablation)DeLo · UPMC-Food101-CMML (eta=70%)AP: 61.95 -> 65.88Improved
add both alignment and consistency losses (ablation)DeLo · UPMC-Food101-CMML (eta=70%)AP: 61.95 -> 67.42Improved
use DeLo instead of RebQRebQ · UPMC-Food101-CMML (eta=90%, image 100%, text 10%)AP: 48.15 -> 67.30Improved
use DeLo instead of RebQRebQ · MM-IMDb-CMML (eta=90%, image 100%, text 10%)AP: 21.37 -> 26.11Improved
use DeLo instead of RebQ (forgetting)RebQ · MM-IMDb-CMML (eta=90%, image 100%, text 10%)FG: 23.66 -> 35.11Hurt

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