Research map / Continual learning / Continual fine-tuning of LLMs
SAME: Stabilized Mixture-of-Experts for Multimodal Continual Instruction Tuning
Zhen-Hao Xie, Jun-Tao Tang, Yu-Cheng Shi, Han-Jia Ye, De-Chuan Zhan, Da-Wei Zhou
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 spectral-aware routing to MoELoRA (ablation) | MoELoRA · CoIN (8 tasks) | average accuracy: 50.58 -> 61.32 | Improved |
| add curvature-aware Riemannian scaling on top of spectral-aware routing (ablation) | SAME · CoIN (8 tasks) | average accuracy: 61.32 -> 65.89 | Improved |
| add adaptive expert activation (freeze selected experts) on top of routing and scaling (ablation) | SAME · CoIN (8 tasks) | average accuracy: 65.89 -> 66.82 | Improved |
| add adaptive expert activation (ablation) | SAME · CoIN (8 tasks) | training time per task: -32.1 min per task, -2.3K MiB/GPU on average | Improved |
| add adaptive expert activation (ablation) | SAME · CoIN OCR-VQA and ScienceQA tasks | per-task accuracy: ScienceQA 80.29 -> 78.35; OCR-VQA 63.19 -> 63.59 | Mixed |
| add curvature-aware scaling on top of routing (ablation) | SAME · CoIN VQAv2 and OCR-VQA tasks | per-task accuracy: VQAv2 64.68 -> 61.21; ScienceQA 71.44 -> 80.29 | Mixed |
| add spectral-aware routing (ablation) | MoELoRA · CoIN Task 1 router overlap after Task 8 | routing overlap (%): 79.40 -> 81.59 | Improved |
| add curvature-aware scaling with router re-trained (ablation) | SAME · CoIN Task 1 recovered accuracy | re-routing accuracy | Improved |
| use SAME instead of MoELoRA | MoELoRA · TriGap (10 tasks) | average accuracy: 44.45 -> 46.53 | Improved |
| use SAME instead of MoELoRA | MoELoRA · ScienceQA case-mismatch error after Task 8 | case mismatch error rate (%): 22.5 -> 7.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- CL-MoE — CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025) · better on 3 of 3 shared results
- SEFE — SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025) · better on 1 of 1 shared result
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
- Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning (ICML 2025 · similarity 0.90)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.89)
- DISTMOE: Rehearsal-free Routing in Mixture-of-Experts for Distributed Instruction Tuning (NeurIPS 2026 · similarity 0.89)
- CoIN: A Benchmark of Continual Instruction Tuning for Multimodel Large Language Models (NeurIPS 2024 · similarity 0.89)
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.88)
- FLAME: Adaptive Mixture-of-Experts for Continual Multimodal Multi-Task Learning (NeurIPS 2026 · similarity 0.87)
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.
