Research map / Continual learning / Continual fine-tuning of LLMs
Mind the Gap: Preserving and Compensating for the Modality Gap in CLIP-Based Continual Learning
Linlan Huang, Xusheng Cao, Haori Lu, Yifan Meng, Fei Yang, Xialei Liu
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 adaptive modality gap preservation (MGP) to simple fine-tuning (ablation) | CLIP ViT-B/16 LoRA · ImageNet100 class-incremental | Last accuracy: 72.74 -> 76.86 | Improved |
| add modality gap compensation (MGC) to simple fine-tuning (ablation) | CLIP ViT-B/16 LoRA · ImageNet100 class-incremental | Last accuracy: 72.74 -> 74.58 | Improved |
| add MGC on top of MGP (ablation) | CLIP ViT-B/16 LoRA · ImageNet100 class-incremental | Last accuracy: 76.86 -> 78.38 | Improved |
| MGP instead of distillation to limit modality gap change (ablation) | CLIP ViT-B/16 · ImageNet100 class-incremental | Last accuracy: 75.85 -> 78.38 | Improved |
| fine-tune naively without gap control (negative similarity falls) (ablation) | CLIP ViT-B/16 · ImageNet100 class-incremental | mean negative cosine similarity: 0.1719 -> 0.0579 | Hurt |
| add 2000 replay samples to MG-CLIP | CLIP ViT-B/16 · CIFAR-100 | Last accuracy: 80.40 -> 82.58 | Improved |
| add 2000 replay samples to MG-CLIP | CLIP ViT-B/16 · ImageNet100 class-incremental | Last accuracy: 78.38 -> 80.74 | Improved |
| vary threshold alpha 5% to 30% (ablation) | CLIP ViT-B/16 · ImageNet100 class-incremental | Last accuracy: 77.88, 78.38, 77.34, 76.9 for 5%, 10%, 20%, 30% | Mixed |
| vary ensemble weight beta 1 to 8 (ablation) | CLIP ViT-B/16 · ImageNet100 class-incremental | Last accuracy: 77.58, 77.92, 78.38, 77.98, 77.32 for 1, 2, 4, 6, 8 | Mixed |
| vary LoRA rank 4 to 32 (ablation) | CLIP ViT-B/16 · ImageNet100 class-incremental | Last accuracy: 78.02, 78.38, 78.32, 78.26 for 4, 8, 16, 32 | No clear effect |
| add EMA (ablation) | CLIP ViT-B/16 · ImageNet-R 10-task | Last accuracy: 82.67 -> 82.92 | No clear effect |
| add per-task epoch estimation (ablation) | CLIP ViT-B/16 · ImageNet-R 10-task | Last accuracy: 82.67 -> 82.82 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MagMax — MagMax: Leveraging Model Merging for Seamless Continual Learning (ECCV 2024) · better on 6 of 6 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
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.90)
- Embracing Language Inclusivity and Diversity in CLIP through Continual Language Learning (AAAI 2024 · similarity 0.89)
- CLAP4CLIP: Continual Learning with Probabilistic Finetuning for Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.87)
- Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models (NeurIPS 2025 · similarity 0.87)
- Bisecle: Binding and Separation in Continual Learning for Video Language Understanding (NeurIPS 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.
