Research map / Continual learning / Continual fine-tuning of LLMs
CTP:Towards Vision-Language Continual Pretraining via Compatible Momentum Contrast and Topology Preservation
Hongguang Zhu, Yunchao Wei, Xiaodan Liang, Chunjie Zhang, Yao Zhao
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 compatible momentum contrast to sequential finetuning (ablation) | ALBEF continual pretraining · P9D cross-modal retrieval | Rm: 56.86 -> 59.86 | Improved |
| add cross-modal topology preservation (ablation) | ALBEF continual pretraining · P9D cross-modal retrieval | Rm: 56.86 -> 62.50 | Improved |
| add cross-modal topology preservation alone (ablation) | ALBEF continual pretraining · P9D multi-modal retrieval | mAP@1: 62.15 -> 61.32 | Hurt |
| add same-modal topology preservation on top of CMC and Lc (ablation) | CTP · P9D cross-modal retrieval | Rm: 64.38 -> 64.87 | Improved |
| use compatible momentum update from both old and current model instead of one side (ablation) | CTP · P9D cross-modal retrieval | Rm: 63.57 -> 64.87 | Improved |
| add momentum queue of negatives (ablation) | CTP · P9D cross-modal retrieval | Rm: 63.94 -> 64.87 | Improved |
| suppress same-modal diagonal similarity (ablation) | CTP · P9D cross-modal retrieval | Rm: 63.12 -> 64.87 | Improved |
| add ER replay buffer (10000 pairs) to CTP (ablation) | CTP · P9D cross-modal retrieval | Rm: 64.87 -> 70.63 | Improved |
| replace parameter regularization (EWC) with CTP | memory-free continual VLP · P9D cross-modal retrieval | Rm: 59.24 -> 64.87 | Improved |
| continual methods on multi-modal retrieval versus sequential finetuning | CTP and baselines · P9D multi-modal retrieval | mAP@1: 62.15 -> 62.64 | No clear effect |
| vary momentum coefficient m from 0.7 to 0.99 (ablation) | CTP · P9D cross-modal retrieval | Rm: 64.83 -> 64.74 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- ICARL — iCaRL: Incremental Classifier and Representation Learning (CVPR 2017) · better on 1 of 2 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
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.91)
- A Practitioner's Guide to Real-World Continual Multimodal Pretraining (NeurIPS 2024 · similarity 0.91)
- Harnessing Textual Semantic Priors for Knowledge Transfer and Refinement in CLIP-Driven Continual Learning (AAAI 2026 · similarity 0.90)
- Dynamic Multi-Layer Null Space Projection for Vision-Language Continual Learning (ICCV 2025 · similarity 0.90)
- CLAP4CLIP: Continual Learning with Probabilistic Finetuning for Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.88)
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.
