Research map / Continual learning / Continual fine-tuning of LLMs
Beyond Disjoint Tasks: Towards More Natural Continual Learning for Vision-Language Models
Xiang Xu, Yiyang Su, Tianchen Zhao, Zheng Zhang, Zhuowen Tu, Anil Jain, Jonathan Wu
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 interference-free delta-space merging instead of factor-level linear merging (ablation) | SOCiaL merge · UCIT-O Protocol I (Tasks 1 and 7) | average accuracy: 65.47 -> 69.85 | Improved |
| add DARE sparsification to delta merge (ablation) | SOCiaL merge · UCIT-O Protocol I (Tasks 1 and 7) | average accuracy: 69.85 -> 69.91 | No clear effect |
| add TIES trimming to delta merge (ablation) | SOCiaL merge · UCIT-O Protocol I (Tasks 1 and 7) | average accuracy: 69.85 -> 68.04 | Hurt |
| add regularized fine-tuning after merge (ablation) | SOCiaL merge · UCIT-O Protocol I (Tasks 1 and 7) | average accuracy: 69.91 -> 86.80 | Improved |
| replace MLP router with bandit router (ablation) | SOCiaL · CoIN | last average task accuracy: 32.74 -> 70.17 | Improved |
| replace mean prototype router with GMM and bandit router (ablation) | SOCiaL · UCIT | last average task accuracy: 69.45 -> 72.23 | Improved |
| raise containment threshold to 0.7 (ablation) | SOCiaL · UCIT-O Protocol I | average accuracy: 64.50 -> 41.79 | Hurt |
| increase GMM components from 1 to 3 (ablation) | SOCiaL · UCIT-O Protocol I | average accuracy: 50.62 -> 64.50 | Improved |
| increase GMM components from 3 to 8 (ablation) | SOCiaL · UCIT-O Protocol I | average accuracy: 64.50 -> 69.41 | Improved |
| increase semantic overlap across protocols | HiDe-LLaVA · UCIT-O | average accuracy: 42.65 -> 34.03 -> 38.76 | Hurt |
| use SOCiaL instead of strongest baseline as overlap grows | SOCiaL · UCIT-O | average accuracy gain: +3.07 on Protocol I to +9.77 on Protocol III | Improved |
| merge overlapping adapters | SOCiaL · UCIT-O Protocol I | number of deployed adapters: 10 -> 6 | Improved |
Closest papers by meaning
- Interference-Aware Continual Vision–Language Learning via Instance-Level Expert Routing (ECCV 2026 · similarity 0.91)
- COMET: Decoupled Distillation, Routing, and Capacity Control for Task-Agnostic Continual Vision--Language Learning (NeurIPS 2026 · similarity 0.91)
- Decomposing and Composing: Towards Efficient Vision-Language Continual Learning via Rank-1 Expert Pool in a Single LoRA (AAAI 2026 · similarity 0.91)
- Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters (CVPR 2024 · similarity 0.90)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.90)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · 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.
