Research map / Continual learning / Continual fine-tuning of LLMs
On Token's Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning of Large Vision Language Models
Chongyang Zhao, Mingsong Li, Haodong Lu, Dong Gong
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 load-balancing auxiliary loss (ablation) | IncMoELoRA · CoIN | MFN: 49.68 -> 50.76 | Improved |
| add Token Assignment Guidance (TAG) (ablation) | IncMoELoRA + Laux · CoIN | MFN: 50.76 -> 54.44 | Improved |
| add Token Assignment Guidance (TAG) (ablation) | IncMoELoRA + Laux · CoIN | BWT: -15.44 -> -7.04 | Improved |
| add exclusivity loss Lexc (ablation) | IncMoELoRA + Laux + TAG · CoIN | MFN: 54.44 -> 55.18 | Improved |
| add specialization loss Lspe (ablation) | IncMoELoRA + Laux + TAG + Lexc · CoIN | MAA: 54.38 -> 57.70 | Improved |
| raise ambiguity threshold from 20% to 50% (ablation) | LLaVA-DyMoE · CoIN | MFN: 57.03 -> 55.32 | Hurt |
| lower ambiguity threshold from 20% to 10% (ablation) | LLaVA-DyMoE · CoIN | MFN: 57.03 -> 56.87 | No clear effect |
| vary loss weight alpha from 1e-3 over 5e-4 to 1e-2 (ablation) | LLaVA-DyMoE · CoIN | MFN: 57.03 -> 55.43 at 1e-2 | No clear effect |
| expand experts per task (IncMoELoRA) instead of static LoRA/MoELoRA | MoELoRA · CoIN | MFN: 43.93 -> 49.68 | Improved |
| add LLaVA-DyMoE to ASD data paradigm instead of SEFE | SEFE · CoIN | BWT: -10.45 -> -4.75 | Improved |
| add LLaVA-DyMoE to ASD data paradigm instead of SEFE | SEFE · CoIN | MAA: 63.04 -> 62.26 | No clear effect |
| combine LLaVA-DyMoE with replay buffer instead of ProgLoRA | ProgLoRA · CoIN | BWT: -6.59 -> -1.55 | Improved |
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.
- SEFE — SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025) · better on 2 of 3 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
- 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)
- CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025 · similarity 0.88)
- Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning (ICML 2025 · similarity 0.88)
- Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model (ICLR 2025 · similarity 0.88)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · 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.
