Research map / Continual learning / Continual fine-tuning of LLMs
Hyperbolic Multimodal Continual Learning
Jiahong Liu, Ming Shen, Xiaohao Liu, Rex Ying, Menglin Yang, Tat-Seng Chua, Irwin King
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 |
|---|---|---|---|
| Use HMCL instead of Vanilla fine-tuning | MERU-L · Continual multimodal classification and retrieval | Overall performance: 38.22 -> 41.36 | Improved |
| Use HMCL instead of Vanilla fine-tuning | MERU-L · Continual multimodal classification and retrieval | Classification BWT: -6.46 -> -0.74 | Improved |
| Use HMCL instead of Vanilla fine-tuning | MERU-B · Continual multimodal classification and retrieval | Overall performance: 41.82 -> 44.21 | Improved |
| Use HMCL instead of Vanilla fine-tuning | HyCoCLIP-B · Continual multimodal classification and retrieval | Retrieval image-to-text R@1: 45.34 -> 57.55 | Improved |
| Use HMCL instead of Vanilla fine-tuning | HyCoCLIP-B · Continual multimodal classification and retrieval | Overall BWT: -5.42 -> -0.43 | Improved |
| Add EWC to Vanilla fine-tuning | MERU-L · Continual multimodal classification and retrieval | Overall performance: 38.22 -> 38.22 | No clear effect |
| Add GEM to Vanilla fine-tuning | MERU-B · Continual multimodal classification and retrieval | Retrieval R@5: 33.66 -> 29.06 | Hurt |
| Add C-FLAT to Vanilla fine-tuning | MERU-L · Continual multimodal classification and retrieval | Overall performance: 38.22 -> 37.16 | Hurt |
| Add C-FLAT to Vanilla fine-tuning | HyCoCLIP-B · Continual multimodal classification and retrieval | Overall performance: 42.33 -> 42.32 | No clear effect |
| Add C-FLAT to Vanilla fine-tuning | MERU-L · Continual multimodal classification and retrieval | Retrieval BWT: -6.46 -> -2.56 | Improved |
| Use HMCL instead of Vanilla on old-task drift | HMCL across datasets · Continual multimodal classification and retrieval | Radial, angular, cross-modal, paired-distance drift: 61.0%, 57.8%, 73.8%, 59.5% reductions | Improved |
| Add block-wise constraint to HMCL updates (ablation) | HMCL · Continual multimodal classification and retrieval | Backward transfer stability | Improved |
Closest papers by meaning
- HypCL: Adapting CLIP in Hyperbolic Space for Continual Learning (ICML 2026 · similarity 0.88)
- Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models (NeurIPS 2025 · similarity 0.88)
- Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training (NeurIPS 2026 · similarity 0.87)
- Accept the Modality Gap: An Exploration in the Hyperbolic Space (CVPR 2024 · similarity 0.87)
- Measuring Representational Shifts in Continual Learning: A Linear Transformation Perspective (ICML 2025 · similarity 0.87)
- Continual Multimodal Contrastive Learning (NeurIPS 2025 · similarity 0.86)
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.
