Research map / Continual learning / Continual fine-tuning of LLMs
Towards Dynamic Modality Alignment in Multimodal Continual Learning
Jiayao Tan, Fan Lyu, Tianle Liu, Fuyuan Hu, Wei Feng
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 graph regularization (all three losses) (ablation) | DAGR · MTIL | Avg. accuracy: 71.2 -> 79.4 | Improved |
| add graph regularization (all three losses) (ablation) | DAGR · MTIL | Last accuracy: 79.5 -> 87.1 | Improved |
| add inter-layer path consistency loss alone (ablation) | DAGR · MTIL | Avg. accuracy: 71.2 -> 75.4 | Improved |
| add multi-hop edge consistency loss alone (ablation) | DAGR · MTIL | Avg. accuracy: 71.2 -> 74.1 | Improved |
| add edge consistency loss alone (ablation) | DAGR · MTIL | Avg. accuracy: 71.2 -> 73.0 | Improved |
| add edge loss to multi-hop plus path losses (ablation) | DAGR · MTIL | Avg. accuracy: 77.6 -> 79.4 | Improved |
| add path loss to edge plus multi-hop losses (ablation) | DAGR · MTIL | Avg. accuracy: 76.2 -> 79.4 | Improved |
| add graph pruning to reduce cost (ablation) | DAGR · MTIL | Avg. accuracy: 79.4 -> 78.8 (time 2.9 h -> 2.6 h) | Hurt |
| add progressive regularization to reduce cost (ablation) | DAGR · MTIL | Avg. accuracy: 79.4 -> 78.6 (time 2.9 h -> 2.7 h) | Hurt |
| add both graph pruning and progressive regularization (ablation) | DAGR · MTIL | Avg. accuracy: 79.4 -> 78.0 (time 2.9 h -> 2.3 h) | Hurt |
| DAGR graph regularization versus DIKI | prompt-based continual learning · MTIL | Avg. accuracy: 76.3 -> 79.4 | Improved |
| DAGR versus DIKI in 16-shot setting | prompt-based continual learning · MTIL-FS (16-shot) | Avg. accuracy: 71.9 -> 74.2 | Improved |
Closest papers by meaning
- Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models (NeurIPS 2025 · similarity 0.89)
- Continual Multimodal Contrastive Learning (NeurIPS 2025 · similarity 0.88)
- Understanding the Emergence of Multimodal Representation Alignment (ICML 2025 · similarity 0.87)
- ReconBoost: Boosting Can Achieve Modality Reconcilement (ICML 2024 · similarity 0.87)
- FLAME: Adaptive Mixture-of-Experts for Continual Multimodal Multi-Task Learning (NeurIPS 2026 · similarity 0.87)
- Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning (ICML 2025 · similarity 0.87)
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.
