Research map / Continual learning / Online and replay-based
Grow-on-Demand: Sparse and Adaptive Expert Expansion for Continual Instruction Tuning
Ying Zhang, Xingyue Guo, Yu Zhao, Xuhui Sui, Baohang Zhou, Xinying Qian, Xiaojie Yuan
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 adaptive Expert Demand Detector instead of always adding an expert (ablation) | GoD-MoE · TRACE (Order-1) | Avg score: 56.1 -> 55.7 | No clear effect |
| use adaptive expansion instead of always adding an expert (ablation) | GoD-MoE · TRACE (Order-1) | trainable parameters: reduces trainable parameters by 35.3% | Improved |
| replace static MoE-LoRA with GoD-MoE adaptive expert expansion and freezing | MoE-LoRA · TRACE (Order-1) | Avg score: 35.6 -> 55.7 | Improved |
| split LoRA into multiple experts (MoE-LoRA) | LoRA · TRACE (Order-1) | Avg score: 23.1 -> 35.6 | Improved |
| add 1% replay data to LoRA | LoRA · TRACE (Order-1) | Avg score: 23.1 -> 50.0 | Improved |
| GoD-MoE without replay versus replay-based PMoE | PMoE · TRACE (Order-1) | Avg score: 51.1 -> 55.7 | Improved |
| GoD-MoE versus replay-based PMoE on forgetting | PMoE · TRACE (Order-1) | BWT: +12.2 -> -3.6 | Hurt |
| GoD-MoE versus second-best baseline on ScienceQA | GoD-MoE · TRACE ScienceQA | score: by 9.3% | Improved |
| GoD-MoE versus second-best baseline on MeetingBank | GoD-MoE · TRACE MeetingBank | score: over 7.7% | Improved |
| change task order from Order-1 to Order-2 and Order-3 (ablation) | GoD-MoE · TRACE | Avg score: 55.7 -> 56.4 and 55.0 | No clear effect |
| continual tuning on TRACE then evaluating GSM | GoD-MoE · GSM | accuracy: 26.1 -> 22.1 | Hurt |
| continual tuning on TRACE then evaluating MMLU | GoD-MoE · MMLU | accuracy: 46.6 -> 48.3 | Improved |
Closest papers by meaning
- CoIN: A Benchmark of Continual Instruction Tuning for Multimodel Large Language Models (NeurIPS 2024 · similarity 0.90)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.89)
- Large Continual Instruction Assistant (ICML 2025 · similarity 0.89)
- Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning (ICML 2025 · similarity 0.88)
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.88)
- Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning (ICLR 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.
