Research map / Continual learning / Continual fine-tuning of LLMs
Mixture-of-experts for multimodal continual tuning
Expands vision-language models with expert adapters and benchmarks continual multimodal learning.
52 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in Continual fine-tuning of LLMs, keyword cluster “unlearning · multimodal · instruction”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
7 papers in 2023–24 and 43 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 13.5% to 22.1%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2022 · CLiMB: A Continual Learning Benchmark for Vision-and-Language Tasks (NeurIPS 2022 · 21 citations)
- 2023 · ConStruct-VL: Data-Free Continual Structured VL Concepts Learning (CVPR 2023 · 13 citations)
- 2024 · Continual Self-supervised Learning: Towards Universal Multi-modal Medical Data Representation Learning (CVPR 2024 · 36 citations)
- 2025 · CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025 · 18 citations)
- 2026 · Drift-aware Collaborative Assistance Mixture of Experts for Heterogeneous Multistream Learning (AAAI 2026 · 4 citations)
Earliest papers here
- CLiMB: A Continual Learning Benchmark for Vision-and-Language Tasks (NeurIPS 2022 · 21 citations)
- BMU-MoCo: Bidirectional Momentum Update for Continual Video-Language Modeling (NeurIPS 2022 · 0 citations)
- ConStruct-VL: Data-Free Continual Structured VL Concepts Learning (CVPR 2023 · 13 citations)
What has been tried
18 papers here report 240 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 5%, hurt 24%, mixed 1%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add feature distillation to ER with random sampling | Average classification tasks · ACC | +3.19% | Helped | Continual Self-supervised Learning: Towards Universal Multi-modal Medical Data Representation Learning (CVPR 2024 · 36 citations) |
| add MMoE to the Vanilla model | VQA v2 (10 tasks) · AP | 32.51 -> 48.43 | Helped | CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025 · 18 citations) |
| Add Rectification with New Knowledge (RNK) | Setting-1, five datasets · average R@1 | 41.2 -> 42.6 | Helped | Continual Vision-Language Retrieval via Dynamic Knowledge Rectification (AAAI 2024 · 13 citations) |
| Initialize the classifier with PLM weights and keep them trainable, rather than freezing them | ImageNet-100 class-IL (B=10, C=10) · Avg accuracy (%) | 57.1% -> 60.3% | Helped | Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · 13 citations) |
| Add the VGA module to the probabilistic encoder | CIFAR100 · Last accuracy | 11.49 -> 73.41 | Helped | CLAP4CLIP: Continual Learning with Probabilistic Finetuning for Vision-Language Models (NeurIPS 2024 · 7 citations) |
| add sparsifying for redundancy reduction (SRR) to LoRA | IconQA · Source average | 44.79 -> 54.73 | Helped | LoRASculpt: Sculpting LoRA for Harmonizing General and Specialized Knowledge in Multimodal Large Language Models (CVPR 2025 · 6 citations) |
| Estimate distributions using all zero-shot predictive probabilities instead of only high-confidence samples and predictive labels | Cross-domain generalization average · top-1 accuracy (%) | 67.22 -> 69.71 | Helped | DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · 3 citations) |
| Increase the old-new weight mixing coefficient | FoMo-in-Flux · AZS | Helped | A Practitioner's Guide to Real-World Continual Multimodal Pretraining (NeurIPS 2024 · 3 citations) | |
| Add replay of a buffer split equally among previous years instead of replaying all past data | TiC-DataComp (M) · Average over 28 datasets | 30.0 -> 28.0 | Hurt | TiC-CLIP: Continual Training of CLIP Models (ICLR 2024 · 1 citation) |
| Add ASD to LoRA training | CoIN · MFN (TA) | 41.59 -> 47.88 | Helped | SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025 · 0 citations) |
| Add knowledge distillation to the V-Adapter | Image generation · FID | 54.61 -> 53.74 | Helped | Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models (NeurIPS 2025 · 0 citations) |
| Add question prediction supervision LQ | NExT-QA · accuracy | 46.58 -> 61.13 | Helped | Bisecle: Binding and Separation in Continual Learning for Video Language Understanding (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 16 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- VQAv2 · 3 of 16 papers
- OCR-VQA · 3 of 16 papers
- GQA · 3 of 16 papers
- TextVQA · 3 of 16 papers
- CoIN · 3 of 16 papers
- ImageNet · 3 of 16 papers
- IconQA · 2 of 16 papers
- ScienceQA · 2 of 16 papers
Baselines
- EWC · 4 of 16 papers
- O-LoRA · 4 of 16 papers
- DER · 3 of 16 papers
- LoRA · 3 of 16 papers
- DualPrompt · 3 of 16 papers
- MoELoRA · 3 of 16 papers
- CL-MoE · 3 of 16 papers
- L2P · 3 of 16 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 50 papers; from OpenAlex and the papers' affiliations).
- Zhejiang University · 2 papers
- Xidian University · 2 papers
- Tencent · 2 papers
- University of New South Wales · 2 papers
Use it
Download BibTeX of 22 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Continual pretraining of language models (66 papers)
- Machine unlearning by selective forgetting (24 papers)
Elsewhere in the topic
- Low-rank and model-merging adaptation (90 papers)
- Synthetic data and merging for CLIP (32 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).