Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Continual Instruction Tuning
Haiyang Guo, Fanhu Zeng, Fei Zhu, Wenzhuo Liu, Da-Han Wang, Jian Xu, Xu-Yao Zhang, Cheng-Lin Liu
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 text-encoder identity tokens instead of image features (ablation) | DISCO · Hom-FCIT task-related beta=1.0 | Last accuracy: 56.22 -> 55.63 without (image only) | Improved |
| use text-only identity tokens instead of text and image combined (ablation) | DISCO · Hom-FCIT task-related beta=1.0 | Last accuracy: 56.22 -> 55.96 without (text & image) | Improved |
| softmax-normalised similarity for activation instead of concatenation (ablation) | DISCO SSA · Hom-FCIT task-related beta=1.0 | Last accuracy: 51.74 -> 56.22 | Improved |
| softmax-normalised similarity instead of raw cosine similarity (ablation) | DISCO SSA · Hom-FCIT task-related beta=1.0 | Last accuracy: 52.83 -> 56.22 | Improved |
| softmax instead of argmax selection of subspace (ablation) | DISCO SSA · Hom-FCIT task-related beta=1.0 | Last accuracy: 55.74 -> 56.22 | Improved |
| softmax instead of concatenation (ablation) | DISCO SSA · Het-FCIT task-related beta=1.0 | Last accuracy: 60.36 -> 63.25 | Improved |
| LoRA in FFN instead of attention only (ablation) | DISCO · Hom-FCIT task-related beta=1.0 | Last accuracy: 56.03 -> 56.22 | Improved |
| LoRA in all linear layers (FFN and attention) instead of FFN only (ablation) | DISCO · Hom-FCIT task-related beta=1.0 | Last accuracy: 56.22 -> 56.30 | No clear effect |
| FedAvg instead of FedAdam aggregation (ablation) | DISCO · Hom-FCIT task-related beta=1.0 | Last accuracy: 55.76 -> 56.22 | Improved |
| FedAvg instead of FedYogi aggregation (ablation) | DISCO · Hom-FCIT task-related beta=1.0 | Avg accuracy: 73.16 vs 73.03 | No clear effect |
| DISCO instead of O-LoRA | Federated continual instruction tuning · Hom-FCIT task-related beta=1.0 | Last accuracy: 49.87 -> 56.22 | Improved |
| DISCO instead of Finetune | Federated continual instruction tuning · Het-FCIT task-related beta=1.0 | Avg accuracy: 54.22 -> 61.99 | Improved |
Closest papers by meaning
- Pilot: Building the Federated Multimodal Instruction Tuning Framework (AAAI 2025 · similarity 0.91)
- Large Continual Instruction Assistant (ICML 2025 · similarity 0.89)
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.89)
- CoIN: A Benchmark of Continual Instruction Tuning for Multimodel Large Language Models (NeurIPS 2024 · similarity 0.88)
- DoFIT: Domain-aware Federated Instruction Tuning with Alleviated Catastrophic Forgetting (NeurIPS 2024 · similarity 0.88)
- Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning (ICML 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.
