Research map / Continual learning / Continual fine-tuning of LLMs
Synthetic data and merging for CLIP
Uses synthetic replay, LoRA recycling and sequential merging to continually adapt vision-language models.
32 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Continual fine-tuning of LLMs, keyword cluster “pre trained · lora · fine tuning”
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.
6 papers in 2023–24 and 25 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 11.5% to 12.8%.
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
- 2021 · Conditionally Adaptive Multi-Task Learning: Improving Transfer Learning in NLP Using Fewer Parameters & Less Data (ICLR 2021 · 39 citations)
- 2023 · Learning to Modulate pre-trained Models in RL (NeurIPS 2023 · 1 citation)
- 2024 · Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · 37 citations)
- 2025 · CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · 23 citations)
- 2026 · ICM-Fusion: In-Context Meta-Optimized LoRA Fusion for Multi-Task Adaptation (AAAI 2026 · 2 citations)
Earliest papers here
- Conditionally Adaptive Multi-Task Learning: Improving Transfer Learning in NLP Using Fewer Parameters & Less Data (ICLR 2021 · 39 citations)
- Learning to Modulate pre-trained Models in RL (NeurIPS 2023 · 1 citation)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · 37 citations)
What has been tried
17 papers here report 204 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 6%, hurt 25%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| use PKT instead of fine-tuning the pre-trained ViT with prototype classifier ψ | CUB200 · ALast accuracy | 3.79±1.47 -> 70.81±0.76 | Helped | Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · 37 citations) |
| add early-exit knowledge distillation (KD) | CIFAR-100 and ImageNet-R · average accuracy | 88.20 -> 90.83; 88.13 -> 90.06; 82.24 -> 83.42; 79.61 -> 80.77 | Helped | CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · 23 citations) |
| add group sparse loss | COCO, forget one class · APf | 0.40 vs. 0 | No clear effect | Continual Forgetting for Pre-trained Vision Models (CVPR 2024 · 15 citations) |
| Add contrastive distillation (CD) to the l2 baseline | MTIL Order I · Avg. | +7.8 | Helped | Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · 5 citations) |
| Add task-specific noise generation and noise mixing | CIFAR-100, 10 steps · average accuracy | 91.18 {{± 0.03}} -> 95.12 {{± 0.05}} | Helped | Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning (NeurIPS 2025 · 3 citations) |
| Vary the projection threshold α | 8, 14, and 20 task sets · ACC | Mixed | Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · 2 citations) | |
| add GPM initialization to PoLoRA-C | ImageNet-R (10 tasks) · average accuracy (%) | 50.33 ± 0.91 -> 62.85 ± 0.61 | Helped | Low-Rank Adaptation in Multilinear Operator Networks for Security-Preserving Incremental Learning (CVPR 2025 · 1 citation) |
| Use changing instead of fixed noise | Safety alignment · NSFW Score | 0.4517 -> 0.4345 | Helped | Leveraging Catastrophic Forgetting to Develop Safe Diffusion Models against Malicious Finetuning (NeurIPS 2024 · 1 citation) |
| Increase reminding-stage steps from 5 to 20 | CIFAR-10 retain set · Accr | 59.8 -> 65.6 | Helped | Targeted Forgetting of Image Subgroups in CLIP Models (CVPR 2025 · 0 citations) |
| replace hard class mining with random class sampling | Few-shot prompt tuning · performance | Hurt | Proxy-FDA: Proxy-based Feature Distribution Alignment for Fine-tuning Vision Foundation Models without Forgetting (ICML 2025 · 0 citations) | |
| Add LADA to BF | X-TAIL 16-shot · Average accuracy | 70.9 -> 72.3 | Helped | LADA: Scalable Label-Specific CLIP Adapter for Continual Learning (ICML 2025 · 0 citations) |
| Replace equivariant zero convolution with Gaussian-initialized standard convolutions | Fine-tuning tasks · performance | Hurt | GeoAda: Efficiently Finetune Geometric Diffusion Models with Equivariant Adapters (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 13 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-100 · 7 of 13 papers
- ImageNet-R · 5 of 13 papers
- EuroSAT · 5 of 13 papers
- ImageNet-A · 4 of 13 papers
- DTD · 4 of 13 papers
- MNIST · 4 of 13 papers
- Caltech101 · 3 of 13 papers
- StanfordCars · 3 of 13 papers
Baselines
- LwF · 5 of 13 papers
- L2P · 4 of 13 papers
- DualPrompt · 3 of 13 papers
- CODA-Prompt · 3 of 13 papers
- EASE · 3 of 13 papers
- WiSE-FT · 3 of 13 papers
- ZSCL · 3 of 13 papers
- APER · 2 of 13 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 31 papers; from OpenAlex and the papers' affiliations).
- Chinese Academy of Sciences · 3 papers
- University of Chinese Academy of Sciences · 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
- Low-rank and model-merging adaptation (90 papers)
Elsewhere in the topic
- Continual pretraining of language models (66 papers)
- Mixture-of-experts for multimodal continual tuning (52 papers)
- Machine unlearning by selective forgetting (24 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).