Research map / Continual learning / Continual fine-tuning of LLMs
TIME: Tensor-Factorized Mixture-of-Experts with Intrinsic Routing for Lifelong Multimodal Knowledge Editing
Dexuan Xu, Jieyi Wang, Shijie Li, Hanpin Wang, Yongzhi Cao, Yu Huang
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 score-threshold expert selection (vs using all experts) (ablation) | TIME · E-VQA lifelong editing (BLIP2, gap=1000) | average score: 90.17 -> 94.15 | Improved |
| add projection-score-based mixing weights (vs average) (ablation) | TIME · E-VQA lifelong editing (BLIP2, gap=1000) | average score: 86.67 -> 94.15 | Improved |
| add the structure-aware alignment loss Lalign (ablation) | TIME · E-VQA lifelong editing (BLIP2, gap=1000) | average score: 93.05 -> 94.15 | Improved |
| use CP decomposition instead of full-rank matrix experts (ablation) | TIME · E-VQA lifelong editing (BLIP2, gap=1000) | average score: 93.49 -> 94.15 | Improved |
| replace LiveEdit's dense experts and external routing with CP experts and intrinsic routing | TIME · E-VQA (BLIP2) | parameter size: 668 MB -> 523 MB | Improved |
| replace LiveEdit's dense experts and external routing with CP experts and intrinsic routing | TIME · E-VQA (LLaVA-v1.5) | parameter size: 1178 MB -> 786 MB | Improved |
| replace LiveEdit's external routing with intrinsic routing | TIME · E-VQA (BLIP2, single editing) | edit time (s): 239.25 -> 163.43 | Improved |
| replace LiveEdit's external routing with intrinsic routing | TIME · E-VQA (BLIP2, gap=1000) | edit time (s): 283.73 -> 178.92 | Improved |
| TIME vs LiveEdit on lifelong editing | TIME · E-VQA (BLIP2, gap=1000) | average score: 93.69 -> 94.15 | Improved |
| TIME vs LiveEdit on lifelong editing | TIME · E-IC (BLIP2, gap=1000) | average score: 82.22 -> 82.53 | Improved |
| TIME vs LiveEdit on single editing | TIME · E-VQA (BLIP2, gap=1) | average score: 96.94 -> 96.70 | No clear effect |
| increase module dimension H from 256 to 2048 (ablation) | TIME · E-VQA lifelong editing (BLIP2, gap=1000) | average score | Improved |
Closest papers by meaning
- Lifelong Knowledge Editing for Vision Language Models with Low-Rank Mixture-of-Experts (CVPR 2025 · similarity 0.90)
- MemEIC: A Step Toward Continual and Compositional Knowledge Editing (NeurIPS 2025 · similarity 0.88)
- WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · similarity 0.87)
- ELDER: Enhancing Lifelong Model Editing with Mixture-of-LoRA (AAAI 2025 · similarity 0.87)
- MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs (NeurIPS 2025 · similarity 0.86)
- Overcoming Generic Knowledge Loss with Selective Parameter Update (CVPR 2024 · 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.
