Research map / Continual learning / Continual fine-tuning of LLMs
SMoLoRA: Exploring and Defying Dual Catastrophic Forgetting in Continual Visual Instruction Tuning
Ziqi Wang, Chang Che, Qi Wang, Yangyang Li, Zenglin Shi, Meng Wang
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add visual understanding (VU) module (ablation) | SeqLoRA · CVIT upstream single-type | AP: 46.21 -> 53.49 | Improved |
| add instruction following (IF) module (ablation) | SeqLoRA · CVIT upstream single-type | AP: 46.21 -> 71.97 | Improved |
| add instruction following (IF) module (ablation) | SeqLoRA · CVIT upstream single-type | MIF: 78.35 -> 98.38 | Improved |
| combine VU and IF by averaging outputs (ablation) | SeqLoRA · CVIT upstream single-type | AP: 71.97 -> 75.16 | Improved |
| add adaptive fusion (AF) module (ablation) | SMoLoRA (VU+IF) · CVIT upstream single-type | AP: 75.16 -> 83.44 | Improved |
| add adaptive fusion (AF) module (ablation) | SMoLoRA (VU+IF) · CVIT upstream single-type | BWT: -10.99 -> -3.23 | Improved |
| add adaptive fusion (AF) module (ablation) | SMoLoRA (VU+IF) · CVIT upstream single-type | MIF: 97.43 -> 97.79 | Improved |
| increase number of LoRA blocks beyond 8 or 16 (ablation) | SMoLoRA · CVIT upstream single-type | all metrics | No clear effect |
| increase number of LoRA blocks from small to 4-8 (ablation) | SMoLoRA · CVIT upstream single-type | all metrics | Improved |
| SMoLoRA versus MoeLoRA on upstream tasks | MoeLoRA · CVIT upstream single-type | AP: 45.27 -> 83.44 | Improved |
| SMoLoRA versus second-best on zero-shot unseen tasks (AP) | DoRA · CVIT zero-shot downstream (single-type) | AP: 25.21 -> 34.35 | Improved |
| SMoLoRA versus second-best few-shot | DoRA · CVIT Places365 few-shot (single-type) | AP: 29.42 -> 40.32 | Improved |
Later papers that compared with this method
1 later accepted paper reports SMoLoRA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- LoRA in LoRA: Towards Parameter-Efficient Architecture Expansion for Continual Visual Instruction Tuning (AAAI 2026 · better on 3 of 3 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
- Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning (NeurIPS 2026 · similarity 0.91)
- Dynamic Cross-Modal Prompt Generation for Multimodal Continual Instruction Tuning (NeurIPS 2026 · similarity 0.90)
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.90)
- CoIN: A Benchmark of Continual Instruction Tuning for Multimodel Large Language Models (NeurIPS 2024 · similarity 0.90)
- SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025 · similarity 0.89)
- Learning to Instruct for Visual Instruction Tuning (NeurIPS 2025 · similarity 0.89)
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.
