Research map / Continual learning
Continual fine-tuning of LLMs: research map
315 papers on Continual fine-tuning of LLMs in Continual learning, from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026), grouped into 2 clusters and 5 approaches. The busiest year so far is 2026.
What it is
Large language and vision-language models are updated again and again with new data and instructions. This topic covers how to add knowledge or skills to such models cheaply (for example with LoRA adapters or model merging) without losing general abilities, and how to remove specific knowledge on request (unlearning).
Terms you will see on this page:
- LoRA
- a cheap fine-tuning method that trains small low-rank matrices next to the frozen weights.
- Model merging
- combining the weights of separately fine-tuned models into one model.
- Machine unlearning
- making a trained model forget specific data or facts.
Start reading with the most cited papers here:
- Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters (CVPR 2024 · 124 citations)
- Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks Using an Incompetent Teacher (AAAI 2023 · 113 citations)
Within Continual learning, its share grew from 14.8% in 2023–24 to 33.9% in 2025–26 (52 → 195 papers at ICML, NeurIPS, ICLR, CVPR and AAAI, the venues with data for all four years).
Explore Continual fine-tuning of LLMs in the interactive map
Working on something in this topic? Describe your idea in scime atlas to see which approach it falls under, the closest papers by meaning and how crowded the spot has become.
What changed recently
Approaches whose share of the topic moved most between 2023–24 and 2025–26 (ICML, NeurIPS, ICLR, CVPR and AAAI):
- Mixture-of-experts for multimodal continual tuning rising: 13% → 22% of the topic (7 → 43 papers)
- Low-rank and model-merging adaptation rising: 33% → 35% of the topic (17 → 69 papers)
- Synthetic data and merging for CLIP rising: 12% → 13% of the topic (6 → 25 papers)
- Continual pretraining of language models falling: 31% → 21% of the topic (16 → 41 papers)
Papers from 2025–2026 that are already cited most:
- CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · 23 citations)
- CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025 · 18 citations)
- Language Guided Concept Bottleneck Models for Interpretable Continual Learning (CVPR 2025 · 10 citations)
- Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · 7 citations)
Approaches and key papers
Continual pretraining, multimodal tuning and unlearning · 160 papers
Keywords: unlearning, multimodal, instruction
Approaches in this cluster:
- Continual pretraining of language models (69 papers)
Analyzes and mitigates forgetting during continual pretraining and knowledge updating of large models. - Mixture-of-experts for multimodal continual tuning (65 papers)
Expands vision-language models with expert adapters and benchmarks continual multimodal learning. - Machine unlearning by selective forgetting (26 papers)
Removes specific data influence from trained networks without full retraining.
Most cited and most cited since 2024:
- Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks Using an Incompetent Teacher (AAAI 2023 · 113 citations)
- LAMOL: LAnguage MOdeling for Lifelong Language Learning (ICLR 2020 · 111 citations)
- Fast Machine Unlearning without Retraining through Selective Synaptic Dampening (AAAI 2024 · 75 citations)
- Continual Self-supervised Learning: Towards Universal Multi-modal Medical Data Representation Learning (CVPR 2024 · 36 citations)
Low-rank adaptation and merging · 155 papers
Keywords: pre trained, lora, fine tuning
Approaches in this cluster:
- Low-rank and model-merging adaptation (114 papers)
Controls forgetting in continual fine-tuning through low-rank updates, prompts and weight merging. - Synthetic data and merging for CLIP (41 papers)
Uses synthetic replay, LoRA recycling and sequential merging to continually adapt vision-language models.
Most cited and most cited since 2024:
- Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters (CVPR 2024 · 124 citations)
- Catastrophic Forgetting Meets Negative Transfer: Batch Spectral Shrinkage for Safe Transfer Learning (NeurIPS 2019 · 101 citations)
- InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning (CVPR 2024 · 69 citations)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · 37 citations)
Related topics in Continual learning
- Online and replay-based (693)
- Domain and prompt-based (138)
- Class-incremental (450)
